A Strong Baseline of Pedestrian Attribute Recognition

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

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]

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
   
   

• 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/
  

• 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
  

• Train baseline based on resnet50

  CUDA_VISIBLE_DEVICES=0 python train.py PETA
  

Reference

• https://github.com/dangweili/pedestrian-attribute-recognition-pytorch
• https://github.com/huanghoujing/EANet