Implementation of Coarse-grained Attribute Learning with Unsupervised Distillation for Real-World Clothes Changing ReID

Paper | Arxiv | Video | More ReID

Table of Contents

• Cite
  
• Setup
  
• ALL Pretrained weights
  
  • Inference & training RLQ (Celeb ReID weights)
  • Logs with pretrained weights will provide configuration details like Batch Size, Seed, etc.
• Inference
  
• Results
  
• Training
  
• Visualization
  

Inference

Analysis has scripts for various evaluations. Mainly RLQ can be evaluated, on datasets like ltcc_cc_gender, prcc_cc_gender, last_cc_gender. Code is dataset independent, just replace dataset argument to do desired inference.

NUM_GPU=2
GPUS=0,1
checkpoint='logs/ltcc_cc_gender/R_LA_15_B=32_1/best_model.pth.tar'
CUDA_VISIBLE_DEVICES=$GPUS python -W ignore -m torch.distributed.launch --nproc_per_node=$NUM_GPU \
  --master_port 12345 main.py --cfg configs/res50_cels_cal_tri_16x4.yaml --dataset ltcc_cc_gender \
  --gpu $GPUS --output ./ --root $ltcc --image --class_2=16 --Pose=$ltcc_pose --pose-mode="R_LA_15" \ --overlap_2=-3 --use_gender $ltcc_gender --extra_class_embed 4096 --extra_class_no 2 --gender_id \
  --backbone="resnet50_joint3_3" --tag output --resume $checkpoint --eval --no-classifier

Results

Results mentioned here are somewhat higher than whats reported in paper. Paper is actually an average of best two runs. Here are providing weights of the best run.

RLQ	Top 1 (CC)	mAP (CC)	Wts & Log
Celeb ReID Base Model	58.1	14.2	Link
Celeb ReID Base Model + CGAL	59.2	14.9	Link
LTCC (Using CelebReID Base Model)	46.4	21.5 / 21.9	Wt1 / Wt2
LTCC (Using CelebReID Base Model + CGAL)	46.7	22.0	Link
LTCC (Using CelebReID Base Model) + 25 Pose Clusters (instead of 15)	46.7	21.7	Link
PRCC (Using CelebReID Base Model)	65.1	63.8	Link
LaST (Using CelebReID Base Model) (4 GPUs)	77.9	35.3	Link
DeepChange (Using CelebReID + Base Model) (6 GPUs)	59.2	22.5	Link

Training

Place Celeb ReID weights in logs/ folder. Update the Celeb_Wt_KL and R_LA_15_2_ABS_GID in scripts such that :

Celeb_Wt_KL=logs/celeb/B=40_KL_4/checkpoint_ep200.pth.tar
R_LA_15_2_ABS_GID=logs/celeb_cc_colors/R_LA_15_2_ABS_GID/best_model.pth.tar

Seeds intialization and Batch size is important. Performance changes a lot across seeds. Thus we recommend running experiments with 1,2,3,4 and reporting an average of best two runs.

Please check Scripts for running various models. We have provided Scripts like : Vanilla CAL model, Base Model, Gender Only, Pose Only, RQL Model.

Most ablation reported in paper is an average of two runs, done on batch size 28 for LTCC and 32 & 40 for PRCC. Best performance for RQL model for LaST, DeepChange is with Batch size 40, and LTCC is on Batch size 40 & 32, and 32 for PRCC.

Code is dataset independent, just replace dataset argument to do desired training.

Pre Processing (Train Only)

All Pose Clusters and Gender related Information for each dataset is kept in Scripts/Helper. This folder also has a list of all RGB images where silhouttes are faulty, and size csv to get a size buckets images fall in.

• (Provided) Genders were manullay generated. (1-> Male / 0 -> female )
• (Provided) Size is measure of spatial resolution.
• (Needed) Silhouttes were generated by Self-Correction for Human Parsing with pretrained weights (checkpoints/exp-schp-201908261155-lip.pth). The script should generate npy dumps.
  • Prep Mask will convert npys to requiste masks (pants and tops)
• (Provided) Alpha Pose is used to dump 2D skeleton. Config: configs/coco/resnet/256x192_res50_lr1e-3_1x.yaml , Model Wt: pretrained_models/fast_res50_256x192.pth
  • Pose clusters are generated via Prep Pose with cluster size of 5,10,15,20,15,25,30,35,40 :

Pose Vector	Description
R_LA	Poses vectors is length and angle of Body lines (Selected for final model)
R_LAC	Poses vectors is length and angle of Body lines + Coordinates of joints
R_A	Poses vectors is only angle of Body lines (Selected for final model)
N_LA	R_LA w/o resizing from 2d skeleton model output (code needs to be modified)
N_LAC	R_LAC w/o resizing from 2d skeleton model output (code needs to be modified)
N_A	R_A w/o resizing from 2d skeleton model output (code needs to be modified)

Visualization

Pose Clusters (Train Only)

Synthetic LQ images (Train Only)

UBD (Train Only)

Citation

If you like our work, please consider citing us:

@inproceedings{Pathak_2025_BMVC,
    author    = {Priyank Pathak and Yogesh S Rawat},
    title     = {Coarse Attribute Prediction with Task Agnostic Distillation for Real World Clothes Changing ReID},
    booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
    publisher = {BMVA},
    year      = {2025},
    url       = {https://bmva-archive.org.uk/bmvc/2025/papers/Paper_346/paper.pdf}
}