The Pytorch Implementation of Category-consistent deep network learning for accurate vehicle logo recognition

• This is the offical website for paper ''Category-consistent deep network learning for accurate vehicle logo recognition''.
• Paper: Category-consistent deep network learning for accurate vehicle logo recognition
• Authors: Wanglong Lu, Hanli Zhao, Qi He, Hui Huang, Xiaogang Jin

Framework Architecture

Requirements

• Pytorch==1.0.1 or higher
• opencv version: 4.1.0

Datasets

• XMU:
  • Y. Huang, R. Wu, Y. Sun, W. Wang, and X. Ding, Vehicle logo recognition system based on convolutional neural networks with a pretraining strategy, IEEE Transactions on Intelligent Transportation Systems 16 (4) (2015) 1951-1960.
  • https://xmu-smartdsp.github.io/VehicleLogoRecognition.html
• HFUT-VL1 and HFUT-VL2:
  • Y. Yu, J. Wang, J. Lu, Y. Xie, and Z. Nie, Vehicle logo recognition based on overlapping enhanced patterns of oriented edge magnitudes, Computers & Electrical Engineering 71 (2018) 273–283.
  • https://github.com/HFUT-VL/HFUT-VL-dataset
• CompCars:
  • L. Yang, P. Luo, C. C. Loy, and X. Tang, A large-scale car dataset for fine-grained categorization and verification, in: Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Boston, MA, 2015, pp. 3973-3981.
  • http://mmlab.ie.cuhk.edu.hk/datasets/comp_cars/index.html
• VLD-45:
  • S. Yang, C. Bo, J. Zhang, P. Gao, Y. Li and S. Serikawa, VLD-45: A big dataset for vehicle logo recognition and detection, IEEE Transactions on Intelligent Transportation Systems (2021) doi: 10.1109/TITS.2021.3062113.
  • https://github.com/YangShuoys/VLD-45-B-DATASET-Detection

VLF-net for classification (Vehicle logo feature extraction network)

• Training with the classification pipeline

  • training XMU dataset

  python train.py --dataset_name XMU --framework Classification_Network
  

  • training HFUT-VL1 dataset

  python train.py --dataset_name HFUT_VL1 --framework Classification_Network
  

  • training HFUT-VL2 dataset

  python train.py --dataset_name HFUT_VL2 --framework Classification_Network
  

  • training CompCars dataset

  python train.py --dataset_name CompCars --framework Classification_Network
  

  • training VLD-45 dataset

  python train.py --dataset_name VLD-45 --framework Classification_Network
  

• Testing with the classification pipeline

  • testing XMU dataset

  python test.py --dataset_name XMU --framework Classification_Network
  

  • testing HFUT-VL1 dataset

  python test.py --dataset_name HFUT_VL1 --framework Classification_Network
  

  • testing HFUT-VL2 dataset

  python test.py --dataset_name HFUT_VL2 --framework Classification_Network
  

  • testing CompCars dataset

  python test.py --dataset_name CompCars --framework Classification_Network
  

  • testing VLD-45 dataset

  python test.py --dataset_name VLD-45 --framework Classification_Network
  

VLF-net for category-consistent mask learning

• Step 1:

  • Generation of the category-consistent masks. There are more details for the co-localization method PSOL.
  • Please note that we use the generated binary-masks directly instead of the predicted boxes.

  1. Download the source codes of [PSOL](https://github.com/tzzcl/PSOL).
  2. Add the generate_mask_imagenet.py from our project to the PSOL folder. 
  3. Run the generate_mask_imagenet.py
  - For the generate_mask_imagenet.py
    --input_size    input size for each image (448 for default)
    --gpu           which gpus to use 
    --output_path   the output pseudo boxes folder
    --batch_size    batch_size for executing forward pass of CNN
    --output_mask_path the output pseudo mask folder
  

• Step 2:

  • After generating the category-consistent masks, we can further organize the training and testing data which are as below:

  root/
        test/
            dog/xxx.png
            dog/xxz.png
            cat/123.png
            cat/nsdf3.png
        train/
            dog/xxx.png
            dog/xxz.png
            cat/123.png
            cat/nsdf3.png
        mask/
            dog/xxx.png
            dog/xxz.png
            cat/123.png
            cat/nsdf3.png
  

  Note that each image has the corresponding generated category-consistent mask.

• Step 3:

  • Now, you can training the model with the category-consistent mask learning framework

  • Training with the category-consistent deep network learning framework pipeline

    • training XMU dataset

    python train.py --dataset_name XMU --framework CCML_Network
    

    • training HFUT-VL1 dataset

    python train.py --dataset_name HFUT_VL1 --framework CCML_Network
    

    • training HFUT-VL2 dataset

    python train.py --dataset_name HFUT_VL2 --framework CCML_Network
    

    • training CompCars dataset

    python train.py --dataset_name CompCars --framework CCML_Network
    

    • training VLD-45 dataset

    python train.py --dataset_name VLD-45 --framework CCML_Network
    

  • Testing with the category-consistent deep network learning framework pipeline

    • testing XMU dataset

    python test.py --dataset_name XMU --framework CCML_Network
    

    • testing HFUT-VL1 dataset

    python test.py --dataset_name HFUT_VL1 --framework CCML_Network
    

    • testing HFUT-VL2 dataset

    python test.py --dataset_name HFUT_VL2 --framework CCML_Network
    

    • testing CompCars dataset

    python test.py --dataset_name CompCars --framework CCML_Network
    

    • testing VLD-45 dataset

    python test.py --dataset_name VLD-45 --framework CCML_Network
    

Experiments

Bibtex

• If you find our code useful, please cite our paper:

  @article{LU2021,
  title = {Category-consistent deep network learning for accurate vehicle logo recognition},
    journal = {Neurocomputing},
    year = {2021},
    issn = {0925-2312},
    doi = {https://doi.org/10.1016/j.neucom.2021.08.030},
    url = {https://www.sciencedirect.com/science/article/pii/S0925231221012145},
    author = {Wanglong Lu and Hanli Zhao and Qi He and Hui Huang and Xiaogang Jin}
    }
  

Acknowledgements

• The authors would like to thank our anonymous reviewers for their valuable comments to improve this paper.
• Our codes benefit from the official memory-efficient-based DenseNet by @gpleiss, Billion-scale-semi-supervised-learning by @leaderj1001, and the implementation of the category-consistent from PSOL by @tzzcl. Thanks for their beautiful work.
• Thanks @kaijieshi7, @Tao Wang, and @Changzu Chen for their help in building the project.