valencebond / Strong_baseline_of_pedestrian_attribute_recognition

A Strong Baseline of Pedestrian Attribute Recognition

The code for our paper Rethinking of Pedestrian Attribute Recognition: Realistic Datasets with Efficient Method.

Considering the big performance gap of various SOTA baseline, we provide a solid and strong baseline for fair comparison.

Update

• 20200901 add infer.py
• 20201215 add peta.mat

Dependencies

• pytorch 1.4.0
• torchvision 0.5.0
• tqdm 4.43.0
• easydict 1.9

Tricks

• sample-wise loss not label-wise loss
• big learning rate combined with clip_grad_norm
• augmentation Pad combined with RandomCrop
• add BN after classifier layer

Performance Comparision

Baseline Performance

• Compared with baseline performance of MsVAA, VAC, ALM, our baseline make a huge performance improvement.
• Compared with our reimplementation of MsVAA, VAC, ALM, our baseline is better.
• We try our best to reimplement MsVAA, VAC and thanks to their code.
• We also try our best to reimplement ALM and try to contact the authors, but no reply received.

SOTA Performance

• Compared with performance of recent state-of-the-art methods, the performance of our baseline is comparable, even better.

• DeepMAR (ACPR15) Multi-attribute Learning for Pedestrian Attribute Recognition in Surveillance Scenarios.
• HPNet (ICCV17) Hydraplus-net: Attentive deep features for pedestrian analysis.
• JRL (ICCV17) Attribute recognition by joint recurrent learning of context and correlation.
• LGNet (BMVC18) Localization guided learning for pedestrian attribute recognition.
• PGDM (ICME18) Pose guided deep model for pedestrian attribute recognition in surveillance scenarios.
• GRL (IJCAI18) Grouping Attribute Recognition for Pedestrian with Joint Recurrent Learning.
• RA (AAAI19) Recurrent attention model for pedestrian attribute recognition.
• VSGR (AAAI19) Visual-semantic graph reasoning for pedestrian attribute recognition.
• VRKD (IJCAI19) Pedestrian Attribute Recognition by Joint Visual-semantic Reasoning and Knowledge Distillation.
• AAP (IJCAI19) Attribute aware pooling for pedestrian attribute recognition.
• MsVAA (ECCV18) Deep imbalanced attribute classification using visual attention aggregation.
• VAC (CVPR19) Visual attention consistency under image transforms for multi-label image classification.
• ALM (ICCV19) Improving Pedestrian Attribute Recognition With Weakly-Supervised Multi-Scale Attribute-Specific Localization.

Dataset Info

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

PA100K[Paper][Github]

RAP : A Richly Annotated Dataset for Pedestrian Attribute Recognition

• v1.0 [Paper][Project]
• v2.0 [Paper][Project]

Zero-shot Protocal

Realistic datasets of PETA and RAPv2 are provided at Google Drive.

You can just replace the 'dataset.pkl' with 'peta_new.pkl' or 'rapv2_new.pkl' to run experiments under new protocal.

Pretrained Models

Pretrained models are provided now at Google Drive.

Because we ran the experiments again, so there may be subtle differences in performance.

Get Started

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

   cd Strong_Baseline_of_Pedestrian_Attribute_Recognition
   mkdir data
   
   

3. Prepare datasets to have following structure:

   ${project_dir}/data
       PETA
           images/
           PETA.mat
           README
       PA100k
           data/
           annotation.mat
           README.txt
       RAP
           RAP_dataset/
           RAP_annotation/
       RAP2
           RAP_dataset/
           RAP_annotation/
   

4. Run the format_xxxx.py to generate dataset.pkl respectively

   python ./dataset/preprocess/format_peta.py
   python ./dataset/preprocess/format_pa100k.py
   python ./dataset/preprocess/format_rap.py
   python ./dataset/preprocess/format_rap2.py
   

5. Train baseline based on resnet50

   CUDA_VISIBLE_DEVICES=0 python train.py PETA
   

Acknowledgements

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

Citation

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

@misc{jia2020rethinking,
    title={Rethinking of Pedestrian Attribute Recognition: Realistic Datasets with Efficient Method},
    author={Jian Jia and Houjing Huang and Wenjie Yang and Xiaotang Chen and Kaiqi Huang},
    year={2020},
    eprint={2005.11909},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}