Scribd
Upload
8 views

Presentation

Uploaded by

Kanute
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
8 views

Presentation

Uploaded by

Kanute
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 37

Bike Price Prediction

Matthias Fast, Amos Dinh


Semester Project
Agenda
1. Problem formulation
2. Crawling for data
3. Filtering and cleaning the data
4. Loss function
5. Model Training
6. Visualization of activations (Conclusion)
1. Problem formulation and motivation

• Suppose you can identify bicycles which are underpriced based on an


image of the bicycle
• You can resell the bicycle to make profit

-> Problem formulation:


„Predict offer-price in euros, given an image of a bike”
2. Crawling
„Predict offer-price in euros, given an image of a bike”
• (Image, price) mapping can be consistently obtained with
webscraping
• scraping of bicycle-images:
• High Quality:
• www.fahrrad.de (1500)
• www.fahrrad-xxl.de (5700)
• Low Quality:
• www.google.com/... &tbm=shop (52000)
~60000 Images
3. Filtering and Cleaning
• Constraints foran im age
• The picture has to contain one bicycle

Reduce the am ountofdata needed (N arrow ed-dow n problem


form ulation):
• The bicycle has to face sidew ays
• The background ofthe im age needs to be m onotonous.

-> Aim to remove deviating images


3. Filtering and Cleaning
3.1. Filter out non-bicycle images
3.2. Remove non-monotonous and other images
3.3. Remove duplicates
3.4. Is there bias in the images?
(3.5.) Preprocessing and mean image
3.1. Filter out non-bicycle images
clf. head embeddings
Chosen method: base model embeddings

1. Compute image embeddings using


ResNet50 CNN (ImageNet)
2. filter images: label “bad samples” by
hand and use their embeddings to
localize similar “bad” images
3. Filter images with visual verification
3.1. Filter out non-bicycle images
3.1. Filter out non-bicycle images
TSNE of ResNet base:
• Embeddings of
classes to be removed
are marked red.
• xnobike (green) is not
well separated
3.1. Filter out non-bicycle images
TSNE of ResNet with
classification head:
• xnobike (green) is
well separated
• other classes (e.g.
blue, cyan) are not
well separated
3.1. Filter out non-bicycle images
• Examination of top-1 predictions of the
classification head images shows many “top-1
classes” to be removable
3.1. Filter out non-bicycle images
• Embeddings after removal, 5200 images are removed,
54000 are kept

-> Still need to remove non-monotonous and xnobike


3.2. Remove non-monotonous and xnobike
1. Perform LDA on embeddings of hand-labeled images vs.
all other images
3.2. Remove non-monotonous and xnobike
2. Manually verify images kept for different thresholds

-> xnobike: Choose threshold 1, remove 1987 images, keep 52021


-> non-monotonous: threshold 0.97, remove 3194 images, keep 48825
3.2. Remove non-monotonous and xnobike
• Removal appears to be mostly successful
3.3. Remove duplicate images
• Many duplicate images (Google)
• Reuse embeddings, since they also represent “similarity”
to a certain degree
• Group images by Euclidean distance, groups of “distance x”

-> Choose threshold 1 , remove 3989 images, keep 44836


3.4. Is there bias in the images?
• Many duplicate images (Google)
3.4. Is there bias in the images?
• Price distribution per source
• The bicycle-sellers tend to have
higher priced bicycles
• Generally, the number of images
from Google outweighs this fact
• Bicycles priced < 100 and > 10000
are removed
3.4. Is there bias in the images?
• Darkness score is the
“average greyscale pixel value”
• Fahrrad.de shows a deviating
distribution but this is not further
pursued
3.4. Is there bias in the images?
• Linear correlation is calculated
• As the price distribution looks
similar to log-normal, this is
explored
• Price is most linearly correlated
with the square root of resolution
and darkness (0.022 and 0.069)
• No large correlation can be
determined
3.5. Preprocessing
• Images are rescaled and
preprocessing applied (for ResNet)
• Cropping of images (Not part of the
NB)
• Visualization immediately before
training
3.5. Mean Image
• Compute the mean image for
different price ranges
• Bicycles are generally right facing
• Most frequent shape seems to be
“Mountainbike” (Non-horizontal
mid-section)
3.5. Datasplit
• Random Shuffle of all 3 sources,
• Google
• Fahrrad.de
• Fahrrad-xxl.de
• Train = 45000
• Dev = 1000
• Test = 1000 (Originally hp. optimization was intended)
• Split is permanently saved and loaded via .json
4. Loss function
• The algorithm will minimize the loss function
• The choice of loss function has to match the problem to be solved
• Otherwise unsatisfactory results will be obtained
4. Loss function
• RMSE:
• MAE:

• MAPE: Mean Absolute Percentage Error


4. Loss function
• RMSE: The tail, high priced image, would
be over-emphasized
• MAPE: „Scaled MAE“ is more sensitive to
small prices
y = 200, ŷ = 300, MAPE = 50%
y = 2000, ŷ = 2100, MAPE = 5%
The original problem: Reselling Profit:
- Allows reselling cheap bicycles as well
- Allows for limited capital inve
5. Model training: Baseline
• The model used is a 50-block ResNet

Regression Head:
• Hidden: 1024 Fully Connected Neurons
• Output: 1 ReLu/Linear Neuron
5. Model training: Baseline
bs: 32
lr: [
0.0001,
0.0003,
0.001
]
epochs: 15

The baseline model can not achieve good performance,


MAPE = 65%
on average: 400€ or 1600€ for a 1000€ bicycle
5. Model training: Conv5-Model
• The model used is a 50-block ResNet

Regression Head:
• Hidden Layer: 1024 Neurons
• Output: 1 ReLu/Linear Neuron
5. Model training: Conv5-Model
bs: 32
lr: 0.0001
epochs: 15

The performance is better, MAPE = 21%


Prediction on average: 800€ or 1200€ for a 1000€ bicycle
Improvements: Hyperparameter Tuning, simply using ResNet V2
5. Model training: Conv5-Model
• The model used is a 50-block ResNet

Retraining:
• The last 6 of the total 53 convolutional layers
Regression Head:
• Output: 1 ReLu/Linear Neuron
5. Model training: Robust Conv5-Model
• The data is augmented by:
• Flip
• Rotation: 2.5 degrees
• MAPE = 22%

• Comparing MAPE

Baseline Conv5 Conv5 aug.


Dev 65% 21% 22%
Aug. Test - 37.9% 22.5%
5. Recap: “Is there bias in the data?”

• Error approximately normal distributed


• No significant error difference between the sources
6. Visualization of activations
• Gradient Cam:
• Gradient of
1. the predicted output with respect to
2. the output-featuremap of
3. each convolutional layer
6. Visualization of activations
• The algorithm generalizes:
• unclean images
• “non-sideview”
images as well
• Grad Cam is not 100%
exact in locality (resolution
is the feature map size of
the respective Conv. Layer)
6. Visualization of activations
• The algorithm picks up on
some random white
background pixelation

-> train-dev-test data-leakage

• Only some images are


affected
• Only some layers are
affected
• The overall impact is not
investigated
Thank you for your attention
Are there any questions, feedback or suggestions?

You might also like