CPI-Parser

Official implementation of CPI-Parser. It based on mmdetection.

Installation

• pytorch 1.10.0
• python 3.7.0
• mmdet 2.25.1

Dataset

You need to download the datasets and annotations follwing this repo's formate

Make sure to put the files as the following structure:

  ├─data
  │  CIHP
  │  │  ├─train_img
  │  │  ├─train_parsing
  │  │  ├─train_seg
  │  │  ├─val_img
  │  │  ├─val_parsing
  │  │  ├─val_seg
  │  │  │─annotations
  │  MHP-v2
  │  │  ├─train_img
  │  │  ├─train_parsing
  │  │  ├─train_seg
  │  │  ├─val_img
  │  │  ├─val_parsing
  │  │  ├─val_seg
  │  │  │─annotations
  |
  ├─work_dirs
  |  ├─causal_parser_r50_inter_8_135k_cihp
  |  |  ├─epoch_75.pth

Results

CIHP

Backbone	LR	mIOU	APvol	AP_p50	PCP50	download
R-50	3x	59.0	58.4	72.2	67.8	model
ConvNext-B	3x	66.5	63.1	80.9	74.6	model

MHP

Backbone	LR	mIOU	APvol	AP_p50	PCP50	download
R-50	3x	38.1	45.8	39.0	49.9	model
ConvNext-B	3x	41.8	48.9	48.6	57.4	model

Evaluation

# inference
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 ./tools/dist_test.sh configs/COSParser/causal_parser_r50_inter_8_135k_cihp.py work_dirs/causal_parser_r50_inter_8_135k_cihp/epoch75.pth 8 --eval bbox --eval-options "jsonfile_prefix=work_dirs/causal_parser_r50_inter_8_135k_cihp/causal_parser_r50_inter_8_135k_cihp_val_result"

# eval, noted that should change the json path produce by previous step.
python utils/eval.py

Training

# training
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 ./tools/dist_train.sh configs/COSParser/causal_parser_r50_inter_8_135k_cihp.py 8