Grab AI Computer Vision Challenge on publicly available "Cars" dataset

DenseNet architecture

      

DenseNet161 architecture was used as a backbone for this project.

Introduction

This repository contains the code in Jypyter Notebook with PyTorch framework pretrained on ImageNet dataset and includes:

For Training:

• Preprocess-Train_Valid.ipynb
• To_SHARPEN_train_images.ipynb
• Train.ipynb
  

For Testing:

• Preprocess-Test.ipynb
• Test.ipynb

Preprocessing (Feature Engineering)

1.Two files from Stanford repository cars_train.tgz and car_devkit.tgz need to be uploaded from repository website. After running Preprocess-Train_Valid.ipynb data folder will be created with train and valid folders respectively with their respective bounding boxes and class names.
2.To make the model more robust some engineering adjustments were made such as Sharpening images by running To_SHARPEN_train_images.ipynb. Sharpening was chosen as original images have noticable number of blur images for almost all car models, so by applying sharpening the model can learn features better.
3.To create the test folder cars_test.tgz need to be uploaded from repository website and Preprocess-Test.ipynb need to be run which will create inside the data folder test images folder.

Training process

Train.ipynb inside the code folder was used to train the model. Training of the model is done using modified data sets which include original and sharpened data with 80/20 partition (overall 14659 images - 13030 for train and 1629 for validation which does not include sharpened data). DenseNet161 model with pretrained ImageNet weights was used as it performed the best among other architectures (https://arxiv.org/pdf/1806.02987.pdf) which achieves almost similar performance as most of the fine-grained models used so far.

Classification results

Network	Top-1 Accuracy		Top-5 Accuracy		Pre-trained model
	Val	Test	Val	Test
DenseNet161	90.67	90.24	98.47	98.43	ImageNet

• The code was totally created from scratch without any reference to above mentioned paper code.
• Different dataset was created in order to increase the accuracy. Sharpening the data which was my own approach indeed made the model more robust.
• The results are obtained by comparing different optimizers and learning rates, which were not described in the paper in details. Adagrad optimizer performed the best among others with 0.001 learning rate.
• Many other fine-grained models were used, however, most of the paper results are impossible to reproduce. I have also tried "Towards Faster Training of Global Covariance Pooling Networks by Iterative Matrix Square Root Normalization" (https://arxiv.org/pdf/1712.01034.pdf), but the model took longer training time with relatively lower accuracy (less than 90%).

Installation and Usage

1. Pytorch 1.0.1 was used
2. type git clone https://github.com/OrkhanHI/Grab-AI-Computer-Vision-Challenge.git
3. pip install -r requirements.txt
4. prepare the dataset as follows

.
├── train
│   ├── 0001 AM General Hummer SUV 2000
│   │   ├── 1sh_00163.jpg
│   │   ├── 1sh_00462.jpg
|   |   └── ...
│   ├── 0002 Acura RL Sedan 2012
│   ├── 0003 Acura TL Sedan 2012
│   ├── ...
│   ├── ...
│   └── 0196 smart fortwo Convertible 2012
└── valid
    ├── 0001 AM General Hummer SUV 2000
    │   ├── 01277.jpg
    │   ├── 01952.jpg
    |   └── ...
    ├── 0002 Acura RL Sedan 2012
    ├── 0003 Acura TL Sedan 2012
    ├── ...
    ├── ...
    └── 0196 smart fortwo Convertible 2012

5. Download best_model_weights from https://drive.google.com/file/d/1vkwNVqTS3ikBNzq4Atz3S7ylBpiTZgX6/view?usp=sharing and copy to outputs folder above (because size of the weights is big please use link to download and copy to the output folder) which will be used in test.ipynb.
6. For Test case preprocess data using Preprocess-Test.ipynb
   Note: Only train images were augmented with sharpen preprocessing. Images with 'sh' prefix mean sharpened, original images do not have any prefixes.

References

Krassimir Valev, Arne Schumann, Lars Sommer, and Jurgen Beyerer,"A Systematic Evaluation of Recent Deep Learning Architectures for Fine-Grained Vehicle Classification", arXiv:1806.02987v1 [cs.CV] 8 Jun 2018

Contact

If you have any questions or suggestions, please contact me

orkhanh.2017@mitb.smu.edu.sg