This paper submitted to TIP is the extension of the previous Arxiv paper.

This project is adopted in the JDAI-CV/fast-reid and PP-Human of PaddleDetection.

1. provide a strong baseline for Pedestrian Attribute Recognition and Multi-Label Classification.
2. provide two new datasets RAPzs and PETAzs following zero-shot pedestrian identity setting.
3. provide a general training pipeline for pedestrian attribute recognition and multi-label classification task.

1. DDP training, which is mainly used for multi-label classifition.
2. Training on all attributes, testing on "selected" attribute. Because the proportion of positive samples for other attributes is less than a threshold, such as 0.01.
   1. For PETA and PETAzs, 35 of the 105 attributes are selected for performance evaluation.
   2. For RAPv1, 51 of the 92 attributes are selected for performance evaluation.
   3. For RAPv2 and RAPzs, 54 and 53 of the 152 attributes are selected for performance evaluation.
   4. For PA100k, all attributes are selected for performance evaluation.
   • However, training on all attributes can not bring consistent performance improvement on various datasets.
3. EMA model.
4. Transformer-base model, such as swin-transformer (with a huge performance improvement) and vit.
5. Convenient dataset info file like dataset_all.pkl

• PETA: Pedestrian Attribute Recognition At Far Distance [Paper][Project]

• PA100K[Paper][Github]

• RAP : A Richly Annotated Dataset for Pedestrian Attribute Recognition

  • v1 [Paper][Project]
  • v2 [Paper][Project]

• PETAzs & RAPzs : Rethinking of Pedestrian Attribute Recognition: A Reliable Evaluation under Zero-Shot Pedestrian Identity Setting Paper [Project]

• The resnet* model is trained by using the weighted function proposed by Tan in AAAI2020.
• Performance in PETAzs and RAPzs based on the first version of PETAzs and RAPzs as described in paper.
• Experiments are conducted on the input size of (256, 192), so there may be minor differences from the results in the paper.
• The reported performance can be achieved at the first drop of learning rate. We also take this model as the best model.
• Pretrained models are provided now at Google Drive.

• python 3.7
• pytorch 1.7.0
• torchvision 0.8.2
• cuda 10.1

1. Run git clone https://github.com/valencebond/Rethinking_of_PAR.git
2. Create a directory to dowload above datasets.

   cd Rethinking_of_PAR
   mkdir data
   

3. Prepare datasets to have following structure:

   ${project_dir}/data
       PETA
           images/
           PETA.mat
           dataset_all.pkl
           dataset_zs_run0.pkl
       PA100k
           data/
           dataset_all.pkl
       RAP
           RAP_dataset/
           RAP_annotation/
           dataset_all.pkl
       RAP2
           RAP_dataset/
           RAP_annotation/
           dataset_zs_run0.pkl
       COCO14
           train2014/
           val2014/
           ml_anno/
               category.json
               coco14_train_anno.pkl
               coco14_val_anno.pkl
   

4. Train baseline based on resnet50

   sh train.sh
   

Codes are based on the repository from Dangwei Li and Houjing Huang. Thanks for their released code.

If you use this method or this code in your research, please cite as:

@article{jia2021rethinking,
  title={Rethinking of Pedestrian Attribute Recognition: A Reliable Evaluation under Zero-Shot Pedestrian Identity Setting},
  author={Jia, Jian and Huang, Houjing and Chen, Xiaotang and Huang, Kaiqi},
  journal={arXiv preprint arXiv:2107.03576},
  year={2021}
}

@inproceedings{jia2021spatial,
 title={Spatial and semantic consistency regularizations for pedestrian attribute recognition},
 author={Jia, Jian and Chen, Xiaotang and Huang, Kaiqi},
 booktitle={Proceedings of the IEEE/CVF international conference on computer vision},
 pages={962--971},
 year={2021}

}