MIC

The document discusses Masked Image Consistency (MIC) for unsupervised domain adaptation. MIC enhances domain adaptation by leveraging spatial context relations between partially and fully visible images. It achieves state-of-the-art results across tasks like semantic segmentation, image classification, and object detection on various datasets.

Uploaded by

mishraaryan846

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PPTX, PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

34 views

MIC

Uploaded by

mishraaryan846

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PPTX, PDF, TXT or read online on Scribd

You are on page 1/ 6

MIC: Masked Image Consistency for Context-Enhanced Domain Adaptation

Abstract
• Unsupervised Domain Adaptation (UDA) adapts models from source to target data without
annotations. Previous methods struggle with visually similar classes in the target domain. The Masked
Image Consistency (MIC) module enhances UDA by leveraging spatial context relations, improving
performance across various tasks and achieving state-of-the-art results in different UDA scenarios.

Presented by Srutanik Bhaduri & Aryan Mishra

3. Methods
For UDA we adjust a model (denoted by θ) to perform well on a

new, unlabeled dataset. Our optimization problem involves two

datasets: the source domain, with NS labeled samples, each

contributing a supervised loss LS_k, and the target domain, with

NT unlabeled samples. Each target sample incurs an adaptation

loss LT_k and a Masked Image Consistency (MIC) loss LM_k.

We use weighting parameters λT and λM to balance these

losses. Our goal is to minimize the total loss across both

domains, ensuring effective adaptation.

Presented by Srutanik Bhaduri & Aryan Mishra

4. Implementation:
4.1 Semantic Segmentation:
• We use datasets like GTA, Synthia, Cityscapes, DarkZurich, and ACDC. They evaluate the MIC approach using

DAFormer and DeepLabV2 networks, starting with pretraining from ImageNet and tweaking training settings like

using AdamW with a learning rate schedule.

4.2 Image Classification:

• We're testing MIC on the VisDA-2017 and Office-Home datasets. We're using ResNet-101 and ViTB/16 models with

a method called SDAT, which combines CDAN with MCC and a smoothness enhancing loss. Our training setup

includes SGD with a learning rate of 0.002, a batch size of 32, and a smoothness parameter of 0.02.

Presented by Srutanik Bhaduri & Aryan Mishra

4. Implementation:

4.3 Object detection:

• In MIC for object detection from Cityscapes to Foggy Cityscapes. We're using Faster R-CNN with ResNet-50 and

FPN, along with a method called SADA for adaptation. The setup follows previous work, including an initial

learning rate of 0.0025 and a batch size of 2. We report results in mean Average Precision (mAP) with a 0.5 IoU

threshold.

5. Conclusion:
Masked Image Consistency (MIC) to improve how models learn from new, unlabeled data. By making sure predictions

match between partially and fully visible images, MIC boosts performance across different tasks and scenarios.

Presented by Srutanik Bhaduri & Aryan Mishra

Explanation of the codebase:
A. The codebase is divided into three repositories:

1. seg - Semantic Segmentation

2. cls - Image Classification

3. det - Object Detection

● We started working on the cls repository which was initially ran on Python 3.8.5 but we ran it on Python
3.8.10 by creating a virtual environment named mic-cls
● We then installed the requirements from the requirements.txt
● We then downloaded two separate datasets and placed them into ‘examples/data/’ after downloading
them separately in ‘/group3/data_cls’:
○ Dataset 1 : Office Dataset
○ Dataset 2: VisDA-2017
● We then run the MIC for Domain Adaptive Classification using the script