About

This is the official implementation of paper Human Parsing Based Alignment with Multi-task Learning for Occluded Person Re-identification, ICME 2020 Oral.

@inproceedings{huang2020human,
  title={Human Parsing Based Alignment With Multi-Task Learning For Occluded Person Re-Identification},
  author={Huang, Houjing and Chen, Xiaotang and Huang, Kaiqi},
  booktitle={2020 IEEE International Conference on Multimedia and Expo (ICME)},
  pages={1--6},
  year={2020},
  organization={IEEE}
}

Requirements

• Python 2.7
• Pytorch 1.0.0
• Torchvision 0.2.1
• No special requirement for sklearn version

Dataset Path

Prepare datasets to have following structure:

• ${project_dir}/dataset
  • coco
    • images
    • masks_7_parts
    • im_name_to_kpt.pkl
    • im_name_to_h_w.pkl
  • market1501
    • Market-1501-v15.09.15
    • Market-1501-v15.09.15_ps_label
  • occ_duke
    • Occluded_Duke
      • bounding_box_train
      • query
      • bounding_box_test
    • Occluded_Duke_ps_label
      • bounding_box_train
      • query
      • bounding_box_test
  • occluded_reid
    • Occluded_REID
      • occluded_body_images
      • whole_body_images
  • partial_reid
    • Partial-REID_Dataset
      • occluded_body_images
      • whole_body_images
  • partial_ilids
    • Partial_iLIDS
      • Probe
      • Gallery
  • partial_ilids_fpr
    • PartialILIDS
      • Probe
      • Gallery

The coco and Market-1501-v15.09.15_ps_label can be downloaded from Baidu Cloud or Google Drive.

To prepare occ_duke,

• First, generate Occluded_Duke following https://github.com/lightas/Occluded-DukeMTMC-Dataset. The result should be placed under ${project_dir}/dataset/occ_duke, with structure
  • ${project_dir}/dataset/occ_duke
    • Occluded_Duke
      • bounding_box_train
      • query
      • bounding_box_test
• Then, download DukeMTMC-reID_ps_label from Baidu Cloud (password g844) or Google Drive. Place it at this location ${project_dir}/occ_duke/DukeMTMC-reID_ps_label.
• Finally, run python script/prepare_occ_duke_ps_label.py.

partial_reid, partial_ilids and partial_ilids_fpr can be downloaded from Baidu Cloud or Google Drive.

Test Baseline

Download model weight from Baidu Cloud (password 21ai) or Google Drive, placing it to location

${project_dir}/exp/train_gp_baseline/market1501/resnet50_tri_lw0.1/ckpt.pth

Then, run the following command

gpu=0 only_test=True bash script/train_gp_baseline.sh

You should get following score (The first line of Table 4 of the paper)

M -> M      [mAP:  72.8%], [cmc1:  90.0%], [cmc5:  95.8%], [cmc10:  97.1%]
M -> PR     [mAP:  54.5%], [cmc1:  53.3%], [cmc5:  75.0%], [cmc10:  86.3%]
M -> PI     [mAP:  53.8%], [cmc1:  48.7%], [cmc5:  71.4%], [cmc10:  79.8%]
M -> PIF    [mAP:  44.4%], [cmc1:  53.2%], [cmc5:  72.9%], [cmc10:  79.8%]
M -> O      [mAP:  54.0%], [cmc1:  62.1%], [cmc5:  79.3%], [cmc10:  85.2%]

Test HPNet

Download model weight from Baidu Cloud (password kv0o) or Google Drive, placing it to location

${project_dir}/exp/train_hpnet/market1501/resnet_sep_conv5_50_tri_lw0.1_seg_lw1_coco_lw1/ckpt.pth

Then, run the following command

gpu=0 only_test=True bash script/train_hpnet.sh

You should get following score (The last lines of Table 3&4 of the paper)

M -> M      [mAP:  74.6%], [cmc1:  91.2%], [cmc5:  97.0%], [cmc10:  98.0%]
M -> PR     [mAP:  81.8%], [cmc1:  85.7%], [cmc5:  95.0%], [cmc10:  96.3%]
M -> PI     [mAP:  72.2%], [cmc1:  68.9%], [cmc5:  82.4%], [cmc10:  89.1%]
M -> PIF    [mAP:  58.9%], [cmc1:  72.0%], [cmc5:  85.3%], [cmc10:  91.3%]
M -> O      [mAP:  77.4%], [cmc1:  87.3%], [cmc5:  93.9%], [cmc10:  96.3%]

Train Baseline

gpu=0 bash script/train_gp_baseline.sh

Train HPNet

gpu=0 bash script/train_hpnet.sh

Ablation study can be found in script/train_hpnet.sh.

Test with Single-gallery-shot Setting

The score in Table 2 of the paper is calculated with single-gallery-shot setting. For this purpose, you can uncomment this line get_scores_str = get_scores_str_r1_r3 in package/eval/eval_feat.py and then set the following three items in config file

cfg.eval.separate_camera_set = True
cfg.eval.single_gallery_shot = True
cfg.eval.first_match_break = False