YU CS Summer 2021 Project | TensorFlow Street Image Classification and Object Detection Model
Download as pptx, pdf0 likes144 views
The document outlines the development of an image classification and object detection model using TensorFlow, intended for self-driving car applications. It details the model's architecture, data sources, training process, and the methodologies employed for object detection, including selective search and non-maximum suppression. The project achieved over 95% accuracy on test data and provides guidance for further enhancements and implementation on GitHub.
1 of 26
Download to read offline
Ad
Recommended
AI powered emotion recognition: From Inception to Production - Global AI Conf...
AI powered emotion recognition: From Inception to Production - Global AI Conf...Vandana Kannan The document provides a comprehensive overview of deep learning, focusing on its principles, training mechanisms, and applications, particularly in emotion recognition. It discusses the differences between deep learning and traditional machine learning, introduces Apache MXNet and Amazon SageMaker as tools for building and deploying models, and outlines essential concepts like CNN architectures, overfitting, and model optimization. Finally, it details the functionalities and features of the MXNet Model Server for serving machine learning models.
AI powered emotion recognition: From Inception to Production - Global AI Conf...
AI powered emotion recognition: From Inception to Production - Global AI Conf...Apache MXNet The document provides a comprehensive overview of deep learning concepts, particularly focusing on the architecture and training of Convolutional Neural Networks (CNNs) using Apache MXNet and Amazon SageMaker. It explains key elements such as different learning types, model training steps, and techniques to prevent overfitting. Additionally, it discusses the deployment and operational aspects of machine learning models using MXNet Model Server and Amazon's associated services.
How Machine Learning Helps Organizations to Work More Efficiently?
How Machine Learning Helps Organizations to Work More Efficiently?Tuan Yang The document discusses machine learning and its application in creating efficient organizations, outlining various clustering strategies and classifiers used for prediction. It emphasizes the importance of understanding the bias-variance trade-off in model generalization and the role of multiple classifier types. Additionally, it highlights the benefits and drawbacks of support vector machines (SVMs) for image classification and suggests that simpler classifiers are often more effective with well-selected features.
Introduction to deep learning
Introduction to deep learningVishwas Lele This document provides an introduction to deep learning. It discusses key concepts such as neural network layers that process input tensors, common layer types like convolutional and recurrent layers, and how networks are trained using stochastic gradient descent. Examples of deep learning applications that have achieved near-human level performance are also presented, such as image classification and speech recognition. The document then focuses on convolutional neural networks, covering concepts like convolution operations, spatial hierarchies, and max pooling. It concludes with a demonstration of digit and X-ray image classification using Keras and techniques for dealing with overfitting like dropout and data augmentation.
Introduction to Deep Learning for Image Analysis at Strata NYC, Sep 2015
Introduction to Deep Learning for Image Analysis at Strata NYC, Sep 2015Turi, Inc. The document provides an overview of deep learning applied to image analysis, covering topics like training models using CIFAR, the mechanics of linear and nonlinear classifiers, and the structure of deep neural networks. It outlines the advantages and challenges of deep learning, including the need for significant computational resources and large datasets, while also discussing the potential of transfer learning to enhance model performance. Ultimately, it emphasizes the effectiveness and versatility of using deep features from pre-trained networks for various tasks in computer vision.
Intro to Deep Learning with Keras - using TensorFlow backend
Intro to Deep Learning with Keras - using TensorFlow backendAmin Golnari This document provides an introduction to convolutional neural networks (CNNs) using the Keras deep learning library. It covers CNN overviews, installing and using Keras, training models on MNIST data, visualizing models, and using the Keras functional API. The document trains basic sequential and convolutional models on MNIST, shows loss/accuracy plots and model layers, and discusses activation functions, optimizers, and the Keras model class API.
Deep Learning Made Easy with Deep Features
Deep Learning Made Easy with Deep FeaturesTuri, Inc. Deep learning models can learn hierarchical feature representations from raw input data. These learned features can then be used to build simple classifiers that achieve high accuracy, even when training data is limited. Transfer learning involves using features extracted from a model pre-trained on a large dataset to build classifiers for other related problems. This approach has been shown to outperform traditional feature engineering with hand-designed features. Deep features extracted from neural networks trained on large image or text datasets have proven to work well as general purpose features for other visual and language problems.
Introduction of Machine learning and Deep Learning
Introduction of Machine learning and Deep LearningMadhu Sanjeevi (Mady) The document provides an overview of machine learning and deep learning, detailing definitions, types of learning (supervised and unsupervised), and the concept of neural networks. It includes information on various algorithms, applications across different domains, and a list of open-source tools available for deep learning development. Key topics discussed include neural network architecture, learning methods, and examples of real-world applications such as spam filtering and recommendation engines.
PyData Delhi 2018 : Creating Art with Neural Nets
PyData Delhi 2018 : Creating Art with Neural Netssrish1 The document discusses neural style transfer using deep learning. It explains that style transfer takes the global structure of one image (the content) and applies the color palettes, textures and local patterns of another image (the style). Deep learning models like convolutional neural networks can be used to separate and recombine these features. The document outlines several papers that developed techniques for neural style transfer, including optimizing loss functions to match content and style features and training feed-forward networks to quickly perform style transfer at runtime.
Deep learning - a primer
Deep learning - a primerUwe Friedrichsen This document is a comprehensive primer on deep learning, addressing its significance, evolution, applications, and underlying principles. It discusses various types of neural networks, including MLP, CNN, RNN, LSTM, and GAN, while emphasizing the importance of training methodologies and challenges like underfitting and overfitting. The content is intended for developers curious about deep learning, highlighting the necessity of mathematical understanding and active engagement in the fast-evolving research landscape.
PR-258: From ImageNet to Image Classification: Contextualizing Progress on Be...
PR-258: From ImageNet to Image Classification: Contextualizing Progress on Be...Jinwon Lee The document discusses issues with the ImageNet dataset and its impact on image classification, highlighting the limitations of data annotation methods and their misalignment with real-world tasks. It identifies biases in label validation and human-label disagreements, particularly in multi-object images, and suggests that existing evaluation metrics may underestimate model performance. The authors propose using human annotators to assess model predictions directly to better understand alignment with human judgment in object classification.
Dato Keynote
Dato KeynoteTuri, Inc. This document discusses the challenges of machine learning development circa 2013 and outlines Dato's approach to addressing these challenges. In 2013, machine learning development was difficult, slow, and expensive. It required specialized knowledge and infrastructure. Dato aims to accelerate the creation of intelligent applications by making sophisticated machine learning as easy as "Hello world" through high-level toolkits, auto feature engineering, automated machine learning (AutoML), and scalable data structures. The document demonstrates how Dato's tools can build an intelligent application with just a few lines of code and handle large datasets by leveraging out-of-core computation.
Handwritten mathematical symbol recognition
Handwritten mathematical symbol recognitionMeghana Kantharaj The document discusses a mathematical symbol recognition system that classifies symbols using machine learning techniques with a training to testing data ratio of 70:30. It outlines the pre-processing steps like binarization and edge detection, along with feature extraction techniques involving zoning and line segment classification, ultimately employing neural networks for symbol classification. Results indicate high efficiency, with 96% accuracy using pixel intensity features and 83% using zoning features.
Object tracking final
Object tracking finalMrsShwetaBanait1 This presentation provides an overview of object tracking, outlining its importance in computer vision and various applications, from security to interface development. It covers definitions, classifications, algorithms, challenges, and future directions for enhancing tracking accuracy through automatic feature selection. The discussion emphasizes the necessity to address issues like occlusion and re-identification in real-time tracking scenarios.
Learn to Build an App to Find Similar Images using Deep Learning- Piotr Teterwak
Learn to Build an App to Find Similar Images using Deep Learning- Piotr TeterwakPyData This document discusses using deep learning and deep features to build an app that finds similar images. It begins with an overview of deep learning and how neural networks can learn complex patterns in data. The document then discusses how pre-trained neural networks can be used as feature extractors for other domains through transfer learning. This reduces data and tuning requirements compared to training new deep learning models. The rest of the document focuses on building an image similarity service using these techniques, including training a model with GraphLab Create and deploying it as a web service with Dato Predictive Services.
Object tracking presentation
Object tracking presentationMrsShwetaBanait1 This presentation by Shweta Kanhere-Banait discusses object tracking in computer vision, including its definition, classifications, algorithms, challenges, applications, and future work. Object tracking is important in various fields such as security and human-computer interaction, while highlighting the roles of different tracking algorithms and their challenges, including appearance changes and occlusions. Future improvements in automatic feature selection and contextual information usage are suggested to enhance tracking accuracy.
Don't Start from Scratch: Transfer Learning for Novel Computer Vision Problem...
Don't Start from Scratch: Transfer Learning for Novel Computer Vision Problem...StampedeCon The document explains transfer learning, which involves adapting a pre-trained neural network to a new dataset, making it especially useful for problems with limited training data. It outlines methods for implementing transfer learning in both classification and detection tasks, emphasizing features such as fine-tuning and feature extraction. Additionally, it provides links to relevant research and further resources for different programming frameworks.
multiple object tracking using particle filter
multiple object tracking using particle filterSRIKANTH DANDE This document presents a project report on multiple object tracking using particle filters. It introduces particle filters and how they are used for object detection and tracking in video surveillance systems. The report reviews three related works that use background subtraction, region-based tracking, and Kalman filters for occlusion handling. It then describes the likelihood function and probability distribution algorithms used in particle filters for reducing unnecessary particles and estimating object locations. The report compares the current work to two papers, noting differences in the use of color vs location information and how particle filters handle occlusion.
Open Source AI and ML, Whats Possible Today?
Open Source AI and ML, Whats Possible Today?Justin Reock The presentation by Justin Reock emphasizes real-world applications of deep learning and artificial intelligence (AI), explaining deep neural networks (DNNs) as complex statistical models that recognize patterns in data. Examples from companies like Airbnb and GE Healthcare illustrate how DNNs have been employed to automate tasks like image classification and enhance medical imaging accuracy. The discussion also touches on the challenges of acquiring labeled data for training AI models and highlights the importance of collaborative open-source efforts in advancing these technologies.
Introduction to machine learning
Introduction to machine learningSanghamitra Deb The document provides an introduction to machine learning, covering its definitions, types (supervised and unsupervised learning), models, and applications across various fields such as healthcare, finance, and education. It also delves into specific algorithms like linear regression, decision trees, and clustering methods, highlighting their use cases, advantages, and challenges, along with performance metrics for evaluating models. Additionally, it discusses techniques for feature engineering and optimization methods that enhance model accuracy and efficiency.
A Large-scale hierarchical image database
A Large-scale hierarchical image databaseRezapourabbas The document introduces ImageNet, a large-scale hierarchical image database that organizes tens of millions of quality-controlled images into synsets based on WordNet concepts. It highlights the database's properties, including scale, hierarchy, accuracy, and diversity while comparing it with other existing datasets. Future work aims to complete and expand ImageNet to make it more accessible and to foster an online community.
Object tracking
Object trackingSri vidhya k The document discusses object tracking in computer vision. It begins with an introduction and overview of applications of object tracking. It then discusses object representation, detection, tracking algorithms and methodologies. It compares different tracking methods and provides an example of object tracking in MATLAB. Key steps in object tracking include object detection, tracking the detected objects across frames using algorithms like point tracking, kernel tracking and silhouette tracking. Common challenges with object tracking are also summarized.
Multiple Object Tracking
Multiple Object TrackingRainakSharma Multiple object tracking (MOT) involves localizing and identifying multiple moving objects over time using video input. MOT has various applications including human-computer interaction, surveillance, and medical imaging. It allows too many detected objects to be matched across frames and tracks objects even if detection fails in some frames. However, challenges include implementing real-time tracking due to batch-based algorithms and solving identity switches and fragmentation when detections are missed. Common MOT methods include Faster R-CNN for detection, Kalman filters for prediction, CNNs for appearance features, and the Hungarian algorithm for data association and tracking.
Advance deep learning
Advance deep learningaliaKhan71 The document provides a comprehensive overview of advanced deep learning techniques, detailing their evolution, various approaches including supervised, unsupervised, and hybrid learning, and popular architectures like CNNs and RNNs. It discusses the methods to improve model performance, including backpropagation, dropout, and transfer learning, as well as the frameworks available for implementing deep learning. Additionally, it highlights various applications of deep learning, emphasizing its potential to enhance diverse fields and contribute positively to humanity.
An introduction to Machine Learning (and a little bit of Deep Learning)
An introduction to Machine Learning (and a little bit of Deep Learning)Thomas da Silva Paula The document introduces the fundamentals of machine learning, describing how a computer program can learn from experience based on certain tasks and performance measures. It covers types of learning, including supervised and unsupervised learning, as well as deep learning techniques such as neural networks and convolutional networks. Additionally, it highlights advancements that have enabled the rise of deep learning, such as improved hardware and new methodologies.
Object tracking
Object trackingchirase44 This document summarizes an article that reviews object tracking methods. It categorizes tracking approaches based on the object and motion representations used. It describes common object representations like points, shapes, contours and appearance models. It also discusses popular image features for tracking like color, motion, edges. The document aims to help readers select suitable tracking algorithms for their applications.
Object tracking a survey
Object tracking a surveyHaseeb Hassan This document is a comprehensive survey of object tracking methodologies, summarizing the evolution and categorization of tracking methods from 1979 to 2006. It discusses challenges in tracking, such as abrupt motion and occlusions, as well as various applications including automated surveillance and traffic monitoring. The paper emphasizes the importance of object representations, feature selection, and the role of detection mechanisms in the tracking process.
Video object tracking with classification and recognition of objects
Video object tracking with classification and recognition of objectsManish Khare The document discusses an ongoing research project on video object tracking using classification and recognition. It presents the initial progress made, including work on automatic image segmentation using level set methods and detection/removal of shadows. Level set methods allow flexible representation of object contours and boundaries during segmentation. The research aims to automatically track and classify multiple objects in video sequences.
cnn ppt.pptx
cnn ppt.pptxrohithprabhas1 Convolutional neural networks (CNNs) are a type of deep neural network commonly used for analyzing visual imagery. CNNs use various techniques like convolution, ReLU activation, and pooling to extract features from images and reduce dimensionality while retaining important information. CNNs are trained end-to-end using backpropagation to update filter weights and minimize output error. Overall CNN architecture involves an input layer, multiple convolutional and pooling layers to extract features, fully connected layers to classify features, and an output layer. CNNs can be implemented using sequential models in Keras by adding layers, compiling with an optimizer and loss function, fitting on training data over epochs with validation monitoring, and evaluating performance on test data.
Deep learning with keras
Deep learning with kerasMOHITKUMAR1379 Keras is an open-source Python framework developed for deep learning that enables fast experimentation with neural networks, supporting both CNN and RNN architectures. It includes functionalities like rapid prototyping, model building (sequential and functional), image data augmentation, and callbacks to enhance training. The framework also offers access to various pre-trained models and datasets, with applications in tasks such as image classification through transfer learning.
More Related Content
What's hot (20)
PyData Delhi 2018 : Creating Art with Neural Nets
PyData Delhi 2018 : Creating Art with Neural Netssrish1 The document discusses neural style transfer using deep learning. It explains that style transfer takes the global structure of one image (the content) and applies the color palettes, textures and local patterns of another image (the style). Deep learning models like convolutional neural networks can be used to separate and recombine these features. The document outlines several papers that developed techniques for neural style transfer, including optimizing loss functions to match content and style features and training feed-forward networks to quickly perform style transfer at runtime.
Deep learning - a primer
Deep learning - a primerUwe Friedrichsen This document is a comprehensive primer on deep learning, addressing its significance, evolution, applications, and underlying principles. It discusses various types of neural networks, including MLP, CNN, RNN, LSTM, and GAN, while emphasizing the importance of training methodologies and challenges like underfitting and overfitting. The content is intended for developers curious about deep learning, highlighting the necessity of mathematical understanding and active engagement in the fast-evolving research landscape.
PR-258: From ImageNet to Image Classification: Contextualizing Progress on Be...
PR-258: From ImageNet to Image Classification: Contextualizing Progress on Be...Jinwon Lee The document discusses issues with the ImageNet dataset and its impact on image classification, highlighting the limitations of data annotation methods and their misalignment with real-world tasks. It identifies biases in label validation and human-label disagreements, particularly in multi-object images, and suggests that existing evaluation metrics may underestimate model performance. The authors propose using human annotators to assess model predictions directly to better understand alignment with human judgment in object classification.
Dato Keynote
Dato KeynoteTuri, Inc. This document discusses the challenges of machine learning development circa 2013 and outlines Dato's approach to addressing these challenges. In 2013, machine learning development was difficult, slow, and expensive. It required specialized knowledge and infrastructure. Dato aims to accelerate the creation of intelligent applications by making sophisticated machine learning as easy as "Hello world" through high-level toolkits, auto feature engineering, automated machine learning (AutoML), and scalable data structures. The document demonstrates how Dato's tools can build an intelligent application with just a few lines of code and handle large datasets by leveraging out-of-core computation.
Handwritten mathematical symbol recognition
Handwritten mathematical symbol recognitionMeghana Kantharaj The document discusses a mathematical symbol recognition system that classifies symbols using machine learning techniques with a training to testing data ratio of 70:30. It outlines the pre-processing steps like binarization and edge detection, along with feature extraction techniques involving zoning and line segment classification, ultimately employing neural networks for symbol classification. Results indicate high efficiency, with 96% accuracy using pixel intensity features and 83% using zoning features.
Object tracking final
Object tracking finalMrsShwetaBanait1 This presentation provides an overview of object tracking, outlining its importance in computer vision and various applications, from security to interface development. It covers definitions, classifications, algorithms, challenges, and future directions for enhancing tracking accuracy through automatic feature selection. The discussion emphasizes the necessity to address issues like occlusion and re-identification in real-time tracking scenarios.
Learn to Build an App to Find Similar Images using Deep Learning- Piotr Teterwak
Learn to Build an App to Find Similar Images using Deep Learning- Piotr TeterwakPyData This document discusses using deep learning and deep features to build an app that finds similar images. It begins with an overview of deep learning and how neural networks can learn complex patterns in data. The document then discusses how pre-trained neural networks can be used as feature extractors for other domains through transfer learning. This reduces data and tuning requirements compared to training new deep learning models. The rest of the document focuses on building an image similarity service using these techniques, including training a model with GraphLab Create and deploying it as a web service with Dato Predictive Services.
Object tracking presentation
Object tracking presentationMrsShwetaBanait1 This presentation by Shweta Kanhere-Banait discusses object tracking in computer vision, including its definition, classifications, algorithms, challenges, applications, and future work. Object tracking is important in various fields such as security and human-computer interaction, while highlighting the roles of different tracking algorithms and their challenges, including appearance changes and occlusions. Future improvements in automatic feature selection and contextual information usage are suggested to enhance tracking accuracy.
Don't Start from Scratch: Transfer Learning for Novel Computer Vision Problem...
Don't Start from Scratch: Transfer Learning for Novel Computer Vision Problem...StampedeCon The document explains transfer learning, which involves adapting a pre-trained neural network to a new dataset, making it especially useful for problems with limited training data. It outlines methods for implementing transfer learning in both classification and detection tasks, emphasizing features such as fine-tuning and feature extraction. Additionally, it provides links to relevant research and further resources for different programming frameworks.
multiple object tracking using particle filter
multiple object tracking using particle filterSRIKANTH DANDE This document presents a project report on multiple object tracking using particle filters. It introduces particle filters and how they are used for object detection and tracking in video surveillance systems. The report reviews three related works that use background subtraction, region-based tracking, and Kalman filters for occlusion handling. It then describes the likelihood function and probability distribution algorithms used in particle filters for reducing unnecessary particles and estimating object locations. The report compares the current work to two papers, noting differences in the use of color vs location information and how particle filters handle occlusion.
Open Source AI and ML, Whats Possible Today?
Open Source AI and ML, Whats Possible Today?Justin Reock The presentation by Justin Reock emphasizes real-world applications of deep learning and artificial intelligence (AI), explaining deep neural networks (DNNs) as complex statistical models that recognize patterns in data. Examples from companies like Airbnb and GE Healthcare illustrate how DNNs have been employed to automate tasks like image classification and enhance medical imaging accuracy. The discussion also touches on the challenges of acquiring labeled data for training AI models and highlights the importance of collaborative open-source efforts in advancing these technologies.
Introduction to machine learning
Introduction to machine learningSanghamitra Deb The document provides an introduction to machine learning, covering its definitions, types (supervised and unsupervised learning), models, and applications across various fields such as healthcare, finance, and education. It also delves into specific algorithms like linear regression, decision trees, and clustering methods, highlighting their use cases, advantages, and challenges, along with performance metrics for evaluating models. Additionally, it discusses techniques for feature engineering and optimization methods that enhance model accuracy and efficiency.
A Large-scale hierarchical image database
A Large-scale hierarchical image databaseRezapourabbas The document introduces ImageNet, a large-scale hierarchical image database that organizes tens of millions of quality-controlled images into synsets based on WordNet concepts. It highlights the database's properties, including scale, hierarchy, accuracy, and diversity while comparing it with other existing datasets. Future work aims to complete and expand ImageNet to make it more accessible and to foster an online community.
Object tracking
Object trackingSri vidhya k The document discusses object tracking in computer vision. It begins with an introduction and overview of applications of object tracking. It then discusses object representation, detection, tracking algorithms and methodologies. It compares different tracking methods and provides an example of object tracking in MATLAB. Key steps in object tracking include object detection, tracking the detected objects across frames using algorithms like point tracking, kernel tracking and silhouette tracking. Common challenges with object tracking are also summarized.
Multiple Object Tracking
Multiple Object TrackingRainakSharma Multiple object tracking (MOT) involves localizing and identifying multiple moving objects over time using video input. MOT has various applications including human-computer interaction, surveillance, and medical imaging. It allows too many detected objects to be matched across frames and tracks objects even if detection fails in some frames. However, challenges include implementing real-time tracking due to batch-based algorithms and solving identity switches and fragmentation when detections are missed. Common MOT methods include Faster R-CNN for detection, Kalman filters for prediction, CNNs for appearance features, and the Hungarian algorithm for data association and tracking.
Advance deep learning
Advance deep learningaliaKhan71 The document provides a comprehensive overview of advanced deep learning techniques, detailing their evolution, various approaches including supervised, unsupervised, and hybrid learning, and popular architectures like CNNs and RNNs. It discusses the methods to improve model performance, including backpropagation, dropout, and transfer learning, as well as the frameworks available for implementing deep learning. Additionally, it highlights various applications of deep learning, emphasizing its potential to enhance diverse fields and contribute positively to humanity.
An introduction to Machine Learning (and a little bit of Deep Learning)
An introduction to Machine Learning (and a little bit of Deep Learning)Thomas da Silva Paula The document introduces the fundamentals of machine learning, describing how a computer program can learn from experience based on certain tasks and performance measures. It covers types of learning, including supervised and unsupervised learning, as well as deep learning techniques such as neural networks and convolutional networks. Additionally, it highlights advancements that have enabled the rise of deep learning, such as improved hardware and new methodologies.
Object tracking
Object trackingchirase44 This document summarizes an article that reviews object tracking methods. It categorizes tracking approaches based on the object and motion representations used. It describes common object representations like points, shapes, contours and appearance models. It also discusses popular image features for tracking like color, motion, edges. The document aims to help readers select suitable tracking algorithms for their applications.
Object tracking a survey
Object tracking a surveyHaseeb Hassan This document is a comprehensive survey of object tracking methodologies, summarizing the evolution and categorization of tracking methods from 1979 to 2006. It discusses challenges in tracking, such as abrupt motion and occlusions, as well as various applications including automated surveillance and traffic monitoring. The paper emphasizes the importance of object representations, feature selection, and the role of detection mechanisms in the tracking process.
Video object tracking with classification and recognition of objects
Video object tracking with classification and recognition of objectsManish Khare The document discusses an ongoing research project on video object tracking using classification and recognition. It presents the initial progress made, including work on automatic image segmentation using level set methods and detection/removal of shadows. Level set methods allow flexible representation of object contours and boundaries during segmentation. The research aims to automatically track and classify multiple objects in video sequences.
Similar to YU CS Summer 2021 Project | TensorFlow Street Image Classification and Object Detection Model (20)
cnn ppt.pptx
cnn ppt.pptxrohithprabhas1 Convolutional neural networks (CNNs) are a type of deep neural network commonly used for analyzing visual imagery. CNNs use various techniques like convolution, ReLU activation, and pooling to extract features from images and reduce dimensionality while retaining important information. CNNs are trained end-to-end using backpropagation to update filter weights and minimize output error. Overall CNN architecture involves an input layer, multiple convolutional and pooling layers to extract features, fully connected layers to classify features, and an output layer. CNNs can be implemented using sequential models in Keras by adding layers, compiling with an optimizer and loss function, fitting on training data over epochs with validation monitoring, and evaluating performance on test data.
Deep learning with keras
Deep learning with kerasMOHITKUMAR1379 Keras is an open-source Python framework developed for deep learning that enables fast experimentation with neural networks, supporting both CNN and RNN architectures. It includes functionalities like rapid prototyping, model building (sequential and functional), image data augmentation, and callbacks to enhance training. The framework also offers access to various pre-trained models and datasets, with applications in tasks such as image classification through transfer learning.
Computer vision-nit-silchar-hackathon
Computer vision-nit-silchar-hackathonAditya Bhattacharya The document introduces various computer vision topics including convolutional neural networks, popular CNN architectures, data augmentation, transfer learning, object detection, neural style transfer, generative adversarial networks, and variational autoencoders. It provides overviews of each topic and discusses concepts such as how convolutions work, common CNN architectures like ResNet and VGG, why data augmentation is important, how transfer learning can utilize pre-trained models, how object detection algorithms like YOLO work, the content and style losses used in neural style transfer, how GANs use generators and discriminators, and how VAEs describe images with probability distributions. The document aims to discuss these topics at a practical level and provide insights through examples.
JRs presentation-few-shot-learning-overview @ AI4Media WP5 workshop
JRs presentation-few-shot-learning-overview @ AI4Media WP5 workshopHannes Fassold The document provides an overview of few-shot learning, highlighting its importance in scenarios where large annotated datasets are impractical to obtain. It categorizes methods into data augmentation, metric learning, and meta-learning, showcasing examples such as delta-encoder, prototypical networks, and model-agnostic meta learning (MAML). The research group's focus is on few-shot object detection within media production, emphasizing the iterative addition of new object classes and the development of a user-friendly interface.
Computer Vision for Beginners
Computer Vision for BeginnersSanghamitra Deb This document provides an overview of computer vision techniques including classification and object detection. It discusses popular deep learning models such as AlexNet, VGGNet, and ResNet that advanced the state-of-the-art in image classification. It also covers applications of computer vision in areas like healthcare, self-driving cars, and education. Additionally, the document reviews concepts like the classification pipeline in PyTorch, data augmentation, and performance metrics for classification and object detection like precision, recall, and mAP.
UNSUPERVISED NEURAL.pptx UNSUPERVISED PPT
UNSUPERVISED NEURAL.pptx UNSUPERVISED PPTSalihaBathool Unsupervised neural networks are a type of artificial neural network designed to learn patterns and structures from unlabeled data. Unlike supervised learning, where models are trained with labeled input-output pairs, unsupervised learning algorithms identify hidden patterns, clusters, and structures in the data without explicit guidance.
Key Characteristics:
No labeled data is required.
Finds hidden patterns and structures in data.
Used for clustering, dimensionality reduction, anomaly detection, and feature learning.
Types of Unsupervised Neural Networks:
Autoencoders (AEs):
Used for dimensionality reduction, denoising, and anomaly detection.
Consists of an encoder (compressing input) and a decoder (reconstructing the input).
Variants include Variational Autoencoders (VAEs) for probabilistic feature learning.
Restricted Boltzmann Machines (RBMs):
Energy-based model with visible and hidden layers.
Used in feature extraction and collaborative filtering (e.g., recommendation systems).
elf-Organizing Maps (SOMs):
Clustering algorithm that organizes data in a low-dimensional grid format.
Used for visualizing high-dimensional data.
Generative Adversarial Networks (GANs) (Partially Unsupervised):
Consist of a generator (creates fake data) and a discriminator (differentiates real from fake).
Used for image generation, data augmentation, and creative AI applications.
Deep Belief Networks (DBNs):
Stacked layers of RBMs trained layer-wise.
Used in feature extraction, speech recognition, and dimensionality reduction.
Applications of Unsupervised Neural Networks:
Anomaly Detection: Identifying fraud, cybersecurity threats, or medical anomalies.
Clustering & Segmentation: Customer segmentation, genetic clustering, and text topic modeling.
Dimensionality Reduction: Principal Component Analysis (PCA)-like applications for high-dimensional data visualization.
Feature Learning: Learning efficient representations for downstream tasks like classification or prediction.
elf-Organizing Maps (SOMs):
Clustering algorithm that organizes data in a low-dimensional grid format.
Used for visualizing high-dimensional data.
Generative Adversarial Networks (GANs) (Partially Unsupervised):
Consist of a generator (creates fake data) and a discriminator (differentiates real from fake).
Used for image generation, data augmentation, and creative AI applications.
Deep Belief Networks (DBNs):
Stacked layers of RBMs trained layer-wise.
Used in feature extraction, speech recognition, and dimensionality reduction.
Applications of Unsupervised Neural Networks:
Anomaly Detection: Identifying fraud, cybersecurity threats, or medical anomalies.
Clustering & Segmentation: Customer segmentation, genetic clustering, and text topic modeling.
Dimensionality Reduction: Principal Component Analysis (PCA)-like applications for high-dimensional data visualization.
Feature Learning: Learning efficient representations for downstream tasks like classification or prediction.
elf-Organizing Map
Makine Öğrenmesi ile Görüntü Tanıma | Image Recognition using Machine Learning
Makine Öğrenmesi ile Görüntü Tanıma | Image Recognition using Machine LearningAli Alkan The document provides an introduction to image processing and recognition using machine learning. It discusses how deep learning uses hierarchical neural networks inspired by the human brain to learn representations of image data without requiring manual feature engineering. Deep learning has been applied successfully to problems like computer vision through convolutional neural networks. The document also describes how KNIME can be used as an open-source platform to visually build and run deep learning models for image processing tasks and integrate with other tools. It highlights several image processing and deep learning nodes available in KNIME.
Introduction to Convolutional Neural Networks (CNNs).pptx
Introduction to Convolutional Neural Networks (CNNs).pptxCHRISEVANS269099 The document provides an overview of Convolutional Neural Networks (CNNs) in the context of computer vision, detailing their structure and functionality, including layers like convolution, pooling, and fully connected. It discusses various CNN architectures such as LeNet, AlexNet, and VGG-16, highlighting their contributions to image classification and recognition tasks. Additionally, it covers applications of CNNs in fields like retail, healthcare, and automotive, emphasizing their revolutionary impact on computer vision.
ppt.pdf
ppt.pdfMohanRaj924804 This document describes a study that used convolutional neural networks (CNNs) for animal classification from images. The study proposed a novel method for animal face classification using CNN features. The CNN model was trained on images to classify animals into different classes. The model achieved over 90% accuracy on the test data. The authors concluded that CNNs are well-suited for image classification tasks like animal classification due to their ability to automatically extract relevant features from images. Future work could involve classifying other objects using this deep learning approach.
Introduction to Generative AI refers to a subset of artificial intelligence
Introduction to Generative AI refers to a subset of artificial intelligenceKongu Engineering College, Perundurai, Erode Generative AI refers to a subset of artificial intelligence that focuses on creating new content, such as images, text, music, and even videos, based on the data it has been trained on. Generative AI models learn patterns from large datasets and use these patterns to generate new content.
Deep Computer Vision - 1.pptx
Deep Computer Vision - 1.pptxJawadHaider36 Deep computer vision uses deep learning and machine learning techniques to build powerful vision systems that can analyze raw visual inputs and understand what objects are present and where they are located. Convolutional neural networks (CNNs) are well-suited for computer vision tasks as they can learn visual features and hierarchies directly from data through operations like convolution, non-linearity, and pooling. CNNs apply filters to extract features, introduce non-linearity, and use pooling to reduce dimensionality while preserving spatial data. This repeating structure allows CNNs to learn increasingly complex features to perform tasks like image classification, object detection, semantic segmentation, and continuous control from raw pixels.
Mnist soln
Mnist solnDanishFaisal4 The document describes building a convolutional neural network (CNN) model to classify handwritten digits from the MNIST dataset. It discusses collecting MNIST image and label data, developing a CNN using Keras with TensorFlow that includes convolutional and pooling layers followed by dense layers, training the model for 10 epochs using 5-fold cross-validation, evaluating the trained model on the test set where it achieved 98.82% accuracy, and concluding the CNN can successfully recognize handwritten digits.
AISF19 - Unleash Computer Vision at the Edge
AISF19 - Unleash Computer Vision at the EdgeBill Liu This document discusses the key drivers enabling computer vision at the edge, including new machine learning approaches, optimized model architectures, hardware innovations, and improved software tools. It describes how machine learning has advanced computer vision by enabling end-to-end learning without predefined features. Edge-optimized models like GoogleNet and ShuffleNet are discussed. The proliferation of cameras, embedded processors, and AI accelerators is enabling computer vision everywhere. Open-source tools like OpenCV and frameworks like TensorFlow are supporting development, along with platforms to speed application creation.
Deep learning summary
Deep learning summaryankit_ppt This document provides an overview of deep learning and common deep learning concepts. It discusses that deep learning uses complex neural networks to determine representations of data, rather than requiring humans to engineer features. It also describes convolutional neural networks and how they are better than fully connected networks for tasks like image recognition. Additionally, it covers transfer learning and how pre-trained models can be adapted to new tasks by retraining final layers, reducing data and computation needs. Common deep learning architectures mentioned include AlexNet, VGG16, Inception and MobileNets.
Introduction to transfer learning,aster way of adapting a neural network by e...
Introduction to transfer learning,aster way of adapting a neural network by e...ShatrughanKumar14 transfer learning, Introduction & motivation
Adapting Neural Networks
Process
Transfer Learning
Transferring the knowledge of one model to perform a new task. Cheaper, faster way of adapting a neural network by exploiting their generalization properties
Deeplearning
Deeplearning Nimrita Koul This document provides an outline and overview of training convolutional neural networks. It discusses update rules like stochastic gradient descent, momentum, and Adam. It also covers techniques like data augmentation, transfer learning, and monitoring the training process. The goal of training a CNN is to optimize its weights and parameters to correctly classify images from the training set by minimizing output error through backpropagation and updating weights.
لموعد الإثنين 03 يناير 2022 143 مبادرة #تواصل_تطوير المحاضرة ال 143 من المباد...
لموعد الإثنين 03 يناير 2022 143 مبادرة #تواصل_تطوير المحاضرة ال 143 من المباد...Egyptian Engineers Association The document provides guidance on entering the IT industry with no prior experience, outlining potential starting positions such as help desk roles and testing, along with necessary certifications. It also discusses various career paths in software development, machine learning techniques, and offers insights into traditional versus deep learning algorithms. Additionally, it highlights the importance of ongoing learning and practical application in IT careers.
Teach a neural network to read handwriting
Teach a neural network to read handwritingVipul Kaushal This document discusses teaching a neural network to read handwritten digits using the MNIST dataset. It uses a deep convolutional neural network with convolutional layers to extract features from images, max pooling to enhance dominant features, flatten and dense layers, and softmax activation. The model is compiled and trained using the Adam optimizer on 60,000 training images over multiple epochs, and is tested on 10,000 testing images to classify handwritten digits. Problems in choosing the architecture and loading the MNIST format dataset were addressed by referring to cited articles and resources.
Deep Learning.pptx
Deep Learning.pptxRamya Nellutla Deep learning is a type of machine learning that uses artificial neural networks with multiple layers to extract higher-level features from data. It can learn complex patterns within data and handle large numbers of inputs and outputs. Deep learning is implemented using deep neural networks with multiple hidden layers that learn representations of data through backpropagation. The goal of deep learning is to develop systems that can perform tasks requiring human intelligence like visual perception and speech recognition.
slide-171212080528.pptx
slide-171212080528.pptxSharanrajK22MMT1003 This document summarizes a project on real-time object detection using computer vision techniques. It discusses using a system that can recognize objects in a video stream from a camera and label them with bounding boxes and labels. It notes that most video surveillance footage is uninteresting unless there are moving objects. The project aims to address this by building an accurate, fast object detection system that can run on resource-constrained devices. It proposes using a hybrid CNN-SVM model trained on a large dataset to recognize objects and discusses the training and detection phases of the system.
Introduction to Generative AI refers to a subset of artificial intelligence
Introduction to Generative AI refers to a subset of artificial intelligenceKongu Engineering College, Perundurai, Erode
لموعد الإثنين 03 يناير 2022 143 مبادرة #تواصل_تطوير المحاضرة ال 143 من المباد...
لموعد الإثنين 03 يناير 2022 143 مبادرة #تواصل_تطوير المحاضرة ال 143 من المباد...Egyptian Engineers Association
Ad
Recently uploaded (20)
Lessons Learned from Developing Secure AI Workflows.pdf
Lessons Learned from Developing Secure AI Workflows.pdfPriyanka Aash Lessons Learned from Developing Secure AI Workflows
FIDO Seminar: Evolving Landscape of Post-Quantum Cryptography.pptx
FIDO Seminar: Evolving Landscape of Post-Quantum Cryptography.pptxFIDO Alliance FIDO Seminar: Evolving Landscape of Post-Quantum Cryptography
"How to survive Black Friday: preparing e-commerce for a peak season", Yurii ...
"How to survive Black Friday: preparing e-commerce for a peak season", Yurii ...Fwdays We will explore how e-commerce projects prepare for the busiest time of the year, which key aspects to focus on, and what to expect. We’ll share our experience in setting up auto-scaling, load balancing, and discuss the loads that Silpo handles, as well as the solutions that help us navigate this season without failures.
War_And_Cyber_3_Years_Of_Struggle_And_Lessons_For_Global_Security.pdf
War_And_Cyber_3_Years_Of_Struggle_And_Lessons_For_Global_Security.pdfbiswajitbanerjee38 Russia is one of the most aggressive nations when it comes to state coordinated cyberattacks — and Ukraine has been at the center of their crosshairs for 3 years. This report, provided the State Service of Special Communications and Information Protection of Ukraine contains an incredible amount of cybersecurity insights, showcasing the coordinated aggressive cyberwarfare campaigns of Russia against Ukraine.
It brings to the forefront that understanding your adversary, especially an aggressive nation state, is important for cyber defense. Knowing their motivations, capabilities, and tactics becomes an advantage when allocating resources for maximum impact.
Intelligence shows Russia is on a cyber rampage, leveraging FSB, SVR, and GRU resources to professionally target Ukraine’s critical infrastructures, military, and international diplomacy support efforts.
The number of total incidents against Ukraine, originating from Russia, has steadily increased from 1350 in 2021 to 4315 in 2024, but the number of actual critical incidents has been managed down from a high of 1048 in 2022 to a mere 59 in 2024 — showcasing how the rapid detection and response to cyberattacks has been impacted by Ukraine’s improved cyber resilience.
Even against a much larger adversary, Ukraine is showcasing outstanding cybersecurity, enabled by strong strategies and sound tactics. There are lessons to learn for any enterprise that could potentially be targeted by aggressive nation states.
Definitely worth the read!
FIDO Seminar: Targeting Trust: The Future of Identity in the Workforce.pptx
FIDO Seminar: Targeting Trust: The Future of Identity in the Workforce.pptxFIDO Alliance FIDO Seminar: Targeting Trust: The Future of Identity in the Workforce
Creating Inclusive Digital Learning with AI: A Smarter, Fairer Future
Creating Inclusive Digital Learning with AI: A Smarter, Fairer FutureImpelsys Inc. Have you ever struggled to read a tiny label on a medicine box or tried to navigate a confusing website? Now imagine if every learning experience felt that way—every single day.
For millions of people living with disabilities, poorly designed content isn’t just frustrating. It’s a barrier to growth. Inclusive learning is about fixing that. And today, AI is helping us build digital learning that’s smarter, kinder, and accessible to everyone.
Accessible learning increases engagement, retention, performance, and inclusivity for everyone. Inclusive design is simply better design.
"Database isolation: how we deal with hundreds of direct connections to the d...
"Database isolation: how we deal with hundreds of direct connections to the d...Fwdays What can go wrong if you allow each service to access the database directly? In a startup, this seems like a quick and easy solution, but as the system scales, problems appear that no one could have guessed.
In my talk, I'll share Solidgate's experience in transforming its architecture: from the chaos of direct connections to a service-based data access model. I will talk about the transition stages, bottlenecks, and how isolation affected infrastructure support. I will honestly show what worked and what didn't. In short, we will analyze the controversy of this talk.
AI vs Human Writing: Can You Tell the Difference?
AI vs Human Writing: Can You Tell the Difference?Shashi Sathyanarayana, Ph.D This slide illustrates a side-by-side comparison between human-written, AI-written, and ambiguous content. It highlights subtle cues that help readers assess authenticity, raising essential questions about the future of communication, trust, and thought leadership in the age of generative AI.
Crypto Super 500 - 14th Report - June2025.pdf
Crypto Super 500 - 14th Report - June2025.pdfStephen Perrenod This OrionX's 14th semi-annual report on the state of the cryptocurrency mining market. The report focuses on Proof-of-Work cryptocurrencies since those use substantial supercomputer power to mint new coins and encode transactions on their blockchains. Only two make the cut this time, Bitcoin with $18 billion of annual economic value produced and Dogecoin with $1 billion. Bitcoin has now reached the Zettascale with typical hash rates of 0.9 Zettahashes per second. Bitcoin is powered by the world's largest decentralized supercomputer in a continuous winner take all lottery incentive network.
A Constitutional Quagmire - Ethical Minefields of AI, Cyber, and Privacy.pdf
A Constitutional Quagmire - Ethical Minefields of AI, Cyber, and Privacy.pdfPriyanka Aash A Constitutional Quagmire - Ethical Minefields of AI, Cyber, and Privacy
Smarter Aviation Data Management: Lessons from Swedavia Airports and Sweco
Smarter Aviation Data Management: Lessons from Swedavia Airports and SwecoSafe Software Managing airport and airspace data is no small task, especially when you’re expected to deliver it in AIXM format without spending a fortune on specialized tools. But what if there was a smarter, more affordable way?
Join us for a behind-the-scenes look at how Sweco partnered with Swedavia, the Swedish airport operator, to solve this challenge using FME and Esri.
Learn how they built automated workflows to manage periodic updates, merge airspace data, and support data extracts – all while meeting strict government reporting requirements to the Civil Aviation Administration of Sweden.
Even better? Swedavia built custom services and applications that use the FME Flow REST API to trigger jobs and retrieve results – streamlining tasks like securing the quality of new surveyor data, creating permdelta and baseline representations in the AIS schema, and generating AIXM extracts from their AIS data.
To conclude, FME expert Dean Hintz will walk through a GeoBorders reading workflow and highlight recent enhancements to FME’s AIXM (Aeronautical Information Exchange Model) processing and interpretation capabilities.
Discover how airports like Swedavia are harnessing the power of FME to simplify aviation data management, and how you can too.
GenAI Opportunities and Challenges - Where 370 Enterprises Are Focusing Now.pdf
GenAI Opportunities and Challenges - Where 370 Enterprises Are Focusing Now.pdfPriyanka Aash GenAI Opportunities and Challenges - Where 370 Enterprises Are Focusing Now
The Future of Technology: 2025-2125 by Saikat Basu.pdf
The Future of Technology: 2025-2125 by Saikat Basu.pdfSaikat Basu A peek into the next 100 years of technology. From Generative AI to Global AI networks to Martian Colonisation to Interstellar exploration to Industrial Nanotechnology to Artificial Consciousness, this is a journey you don't want to miss. Which ones excite you the most? Which ones are you apprehensive about? Feel free to comment! Let the conversation begin!
Security Tips for Enterprise Azure Solutions
Security Tips for Enterprise Azure SolutionsMichele Leroux Bustamante Delivering solutions to Azure may involve a variety of architecture patterns involving your applications, APIs data and associated Azure resources that comprise the solution. This session will use reference architectures to illustrate the security considerations to protect your Azure resources and data, how to achieve Zero Trust, and why it matters. Topics covered will include specific security recommendations for types Azure resources and related network security practices. The goal is to give you a breadth of understanding as to typical security requirements to meet compliance and security controls in an enterprise solution.
10 Key Challenges for AI within the EU Data Protection Framework.pdf
10 Key Challenges for AI within the EU Data Protection Framework.pdfPriyanka Aash 10 Key Challenges for AI within the EU Data Protection Framework
Oh, the Possibilities - Balancing Innovation and Risk with Generative AI.pdf
Oh, the Possibilities - Balancing Innovation and Risk with Generative AI.pdfPriyanka Aash Oh, the Possibilities - Balancing Innovation and Risk with Generative AI
2025_06_18 - OpenMetadata Community Meeting.pdf
2025_06_18 - OpenMetadata Community Meeting.pdfOpenMetadata The community meetup was held Wednesday June 18, 2025 @ 9:00 AM PST.
Catch the next OpenMetadata Community Meetup @ https://www.meetup.com/openmetadata-meetup-group/
In this month's OpenMetadata Community Meetup, "Enforcing Quality & SLAs with OpenMetadata Data Contracts," we covered data contracts, why they matter, and how to implement them in OpenMetadata to increase the quality of your data assets!
Agenda Highlights:
👋 Introducing Data Contracts: An agreement between data producers and consumers
📝 Data Contracts key components: Understanding a contract and its purpose
🧑🎨 Writing your first contract: How to create your own contracts in OpenMetadata
🦾 An OpenMetadata MCP Server update!
➕ And More!
CapCut Pro Crack For PC Latest Version {Fully Unlocked} 2025
CapCut Pro Crack For PC Latest Version {Fully Unlocked} 2025pcprocore 👉𝗡𝗼𝘁𝗲:𝗖𝗼𝗽𝘆 𝗹𝗶𝗻𝗸 & 𝗽𝗮𝘀𝘁𝗲 𝗶𝗻𝘁𝗼 𝗚𝗼𝗼𝗴𝗹𝗲 𝗻𝗲𝘄 𝘁𝗮𝗯> https://pcprocore.com/ 👈◀
CapCut Pro Crack is a powerful tool that has taken the digital world by storm, offering users a fully unlocked experience that unleashes their creativity. With its user-friendly interface and advanced features, it’s no wonder why aspiring videographers are turning to this software for their projects.
Raman Bhaumik - Passionate Tech Enthusiast
Raman Bhaumik - Passionate Tech EnthusiastRaman Bhaumik A Junior Software Developer with a flair for innovation, Raman Bhaumik excels in delivering scalable web solutions. With three years of experience and a solid foundation in Java, Python, JavaScript, and SQL, she has streamlined task tracking by 20% and improved application stability.
Ad
YU CS Summer 2021 Project | TensorFlow Street Image Classification and Object Detection Model
- 1. Object Detection in TensorFlow Nissim Cantor, Avi Radinsky, Jacob Silbiger GitHub: https://github.com/ndcantor/tensorflow-street-classifier
- 2. Our Mentor, Gershom Kutliroff • Chief Science Officer, Taranis • CTO & Founder, ClearVuze • Principle Engineer, Intel • CTO & Founder, Omek Interactive • Chief Scientist, IDT Video Technologies Linkedin
- 3. Goal • The goal of our project was to create an image classification and object detection model that simulates how one would be used in a self driving car • Our model uses a neural network to draw boxes around and label objects in an image • Brief overview of model architecture: • TensorFlow transfer learning image classifier • Object detection using: • Selective search • Non-maximum suppression
- 4. Background
- 5. Image Classification vs. Object Detection • Image classification - determining the classification of a given image (ex. Is this a photograph of a cat or a dog). • Object detection - detecting the an image of a certain class within a larger picture (ex. Within this picture of many animals, draw a bounding box around all relevant animals and identify them.)
- 6. Neural Networks • “A neural network is a series of algorithms that endeavors to recognize underlying relationships in a set of data through a process that mimics the way the human brain operates. In this sense, neural networks refer to systems of neurons, either organic or artificial in nature.” (investopedia.com) • It turns out that neural networks often produce generalizable results. • Its best to view them as a black box.
- 7. Convolutional Neural Networks • A class of Neural Networks commonly used in computer vision applications Wikipedia
- 8. VGG16 (Simonyan and Zisserman) • A CNN architecture developed by Oxford University researchers that showed high rates of success in image recognition. • It showed that ‘deep’ Neural Networks might be more effective than ‘large’ neural networks. Wiki
- 9. But how does a Neural Network actually work? • Currently, the prevalent theory as to how a neural network works is that each layer is detecting a specific characteristic of the image that it sees, and as you travel through the layers of the network the layers begin detecting increasingly complex characteristics.
- 10. Transfer Learning • The characteristic detecting nature of neural networks allows us to leverage pre-trained networks on different problems with minimal changes. • The changes can be as minimal as retraining the last layer in the network but sometime can involve training several or more of the last layers in a network.
- 11. Fine Tuning Through tweaking the parameters of a neural network, one can try and maximize its accuracy. However, each tweak of the parameters has its positives and negatives. • Running through the data numerous times (multiple epochs) • It can make up for small data sizes (+) • It can cause overtraining in the model (-) • Increasing the size of the neural network • Accuracy can increase (+) • It becomes easier to overtrain the model (-)
- 13. Data • Since we wanted to simulate a model that might be used in a self driving car, we decided to train our network to recognize everyday objects found in city streets • Classes: • Bicycle • Car • Motorcycle • Person • Train
- 14. Datasets • We used the COCO (Common Objects in Context) Dataset as it contains high- quality images of these objects in the real-life scenes. • We also used Open Images V6 Dataset to supplement some more images for the small sized classes. • Both datasets contained labels and bounding boxes which we used as ground truth to train our model.
- 15. Generating Train/Test Images • All bounding boxes from every image in the datasets were cropped and saved inside of its class folder (crops of cars were stored in the ‘car’ folder, etc) • These cropped images were then split into train and test folders • Shift augmentations were performed for crops in the train folder in order to help the model learn more about the features of each class. • These crops were then used to train the model • Any image from the datasets which had no bounding boxes of the classes we were training the machine to recognize were sliced into quadrants, with each quarter being placed into a ‘background’ class folder • The purpose of the background class is to allow the model to classify unknown objects as ‘background’ instead of mistakenly classifying it as one of the five other classes. • In total the model was trained on over 848,000 cropped images
- 16. Classification Model • Our model uses transfer learning using the MobileNetV2 architecture • The model ran for 20 epochs, the first 10 with a learning rate of 0.0001 and the last 10 with a learning rate of 0.00001 • After training, our model obtained a score of over 95% test accuracy • The figure to the right is a confusion matrix, showing how accurate the model was for each of the 6 classes
- 17. Object Detection • After classifying each image, the next step is object detection • Object detection works by sending image crops into the classification model and drawing boxes around all crops that were classified as one of the non-background classes • We tried two different methods of generating crops to send to into the classifier: 1. Sliding windows 2. Selective search
- 18. Sliding Windows • Sliding windows works by passing a fixed-size rectangle across the image, cropping the image and saving the crop’s location on the original image. The image is then enlarged and the process is repeated in order to generate smaller boxes to help find smaller objects. • The process of repeatedly enlarging the image is called pyramid scaling • The crops are then sent to the classification model to be classified. The model outputs the original image with bounding boxes drawn around all crops which were classified as non- background objects
- 19. Example Original Image Image crops with pyramid scaling
- 20. Problems with Sliding Windows The sliding windows method had two main problems: 1. It is a brute force method, so it took a long time to run 2. The boxes all had a fixed size, so it was hard to classify multiple classes of objects that had different shapes
- 21. Selective Search • Selective search is an algorithm that works by creating crops based on certain criteria in the image • color, texture, shape, size • This produces boxes of different shapes and sizes, and doesn’t work by repeatedly running brute-force across the image • Crops are then sent to the model to be classified Our model uses selective search as it is faster and more accurate than sliding windows (Diagram from pyimagesearch.com)
- 22. Non-maximum Suppression (NMS) • After the cropped regions are classified, many adjacent regions will be classified as the same object and will overlap with each other. In order to choose only the best crop for each object, we run a non-maximum suppression algorithm to remove all overlapping regions. • How NMS works: • Select the crop whose classification the model is most confident about • Remove all bounding boxes whose overlap whose intersection over union (IOU) with the selected crop is above a certain threshold • Output all remaining bounding boxes which are less than the overlap threshold
- 23. Before NMS After NMS
- 24. Results • Our model is able to classify images with relatively high accuracy • We were successful in creating an image detection model that can detect if an object of any of our classes is in the image • With our GitHub repo, you can download the data and train your own model
- 25. Further Exploration • How we might be able to further improve our model: • Uneven training data class sizes led our model to be skewed toward the bigger classes. One way to prevent this would be to find more data to increase the size of the smaller classes • Another way to help this problem is to use a focal loss function. This loss function helps the model adjust its weights more evenly by more drastically adjusting weights when it mistakenly classifies an object of a smaller class than an object of a bigger class • Another aspect to improve is the accuracy of the bounding boxes. In order to have tighter boxes, we would be able to run a bounding box regression algorithm
- 26. Do it Yourself - How to Build and Run Our Model Pre-requisites: 1. Python 3.7+ 2. CUDA 10.0 3. cuDNN 7.5.1 Run: 1. To clone the repo - git clone https://github.com/ndcantor/tensorflow-street-classifier.git 2. Install required Python libraries - pip3 install -r requirements.txt 3. Download data, build, train, and run inference on the model - python3 street_classifier.py 4. Results: A. The model’s train, validation, and text accuracies will be printed on the screen B. A confusion matrix as well as sample test images will be saved to disk