Home

Awesome

Domain-Adaptive 2D Human Pose Estimation via Dual Teachers in Extremely Low-Light Conditions

overview

Introduction

This is the official repository for the ECCV'24 paper: Domain-Adaptive 2D Human Pose Estimation via Dual Teachers in Extremely Low-Light Conditions.

Environment

The code is developed using python 3.11.8 on Ubuntu 20.04, and our model is trained on four NVIDIA RTX 3090 GPUs. Other platforms have not been fully tested.

Quick start

Installation

  1. Clone this repo, and we'll call the directory you cloned as ${POSE_ROOT}.
  2. Install dependencies:
   pip install -r requirements.txt
  1. Install CrowdPoseAPI. Different environments may affect the installation of the API. In our case, we use the following command. (Note: You may encouter an AttributeError when using CrowdPoseAPI due to deprecated numpy features, just follow the prompted instructions and make changes should be fine.)
   pip install -e .
  1. Create output (training model output directory) and log (tensorboard log directory) directories:
   mkdir output 
   mkdir log
  1. Download our pretrained models from (GoogleDrive) and make model directory look like this:
    ${POSE_ROOT}
    |-- model
    `-- |-- imagenet
        |   `-- hrnet_w32-36af842e.pth
         `--exlpose  
            `-- model_final.pth

Directory Tree

Please organize your project directory tree as follows:

   ${POSE_ROOT}
   ├── model
   ├── experiments
   ├── lib
   ├── tools 
   ├── output
   ├── README.md
   └── requirements.txt

Data preparation

For ExLPose data, please download from ExLPose download. Extract them under {DATASET_ROOT}, and make them look like this:

   ${DATASET_ROOT}
   |-- Annotations
   |   |-- ExLPose_test_LL-A.json
   |   |-- ExLPose_test_LL-E.json
   |   |-- ...
   |    `-- ExLPose_train_WL.json
   |-- ExLPose
   |   |-- bright
   |   |   |-- imgs_0119_3_vid000002_exp100_bright_000052__gain_0.00_exposure_1000.png
   |   |   |-- ...
   |   |   `-- imgs_0212_hwangridan_vid000021_exp1200_bright_000092__gain_28.18_exposure_417.png
   |    `-- dark
   |        |-- imgs_0119_3_vid000002_exp100_dark_000052__gain_0.00_exposure_1000.png
   |        |-- ...
   |        `-- imgs_0212_hwangridan_vid000021_exp1200_dark_000092__gain_6.60_exposure_417.png
   `-- ExLPose-OCN
       |-- A7M3
       |    |-- 0822_DSC07102.JPG
       |    |-- ...
       |    `-- 0829_DSC08058.JPG
       `-- RICOH3
            |-- 0825_R0000208.JPG
            |-- ...
            `-- 0829_R0000662.JPG

Testing

Note that the default testing configuration uses 4 GPUs. Please adjust this according to your machine’s specifications.

Testing on ExLPose-test LL-A split

   python tools/valid_test.py --cfg experiments/exlpose/test_config.yaml \
       TEST.MODEL_FILE model/exlpose/model_final.pth DATASET.ROOT ${DATASET_ROOT} \
            TEST.NMS_THRE 0.15 TEST.SCALE_FACTOR 0.5,1,2 TEST.MATCH_HMP True DATASET.TEST all

Set DATASET.TEST to 'normal', 'hard' or 'extreme' to evaluate on LL-N, LL-H and LL-E splits.

Testing on ExLPose-OCN RICOH3 split

   python tools/valid_ocn.py --cfg experiments/exlpose/test_config.yaml \
       TEST.MODEL_FILE model/exlpose/model_final.pth DATASET.ROOT ${DATASET_ROOT} \
           TEST.NMS_THRE 0.15 TEST.SCALE_FACTOR 0.5,1,2 TEST.MATCH_HMP True DATASET.TEST RICOH3

Set DATASET.TEST to 'A7M3' to evaluate on A7M3 splits.

Training on ExLPose

Pre-Training Stage: Training the main teacher on well-lit image

   python tools/train_stage1_PT.py --cfg experiments/exlpose/PT_stage_config.yaml \
       DATASET.ROOT ${DATASET_ROOT}

Pre-Training Stage: Training the main teacher on fake low-light image

   python tools/train_stage1_PT.py --cfg experiments/exlpose/PT_stage_config.yaml \
       DATASET.ROOT ${DATASET_ROOT} TRAIN.STAGE PT_LL MODEL.PRETRAINED_MAIN ${MAIN_WEIGHTS_FILE}

Pre-Training Stage: Training the comp. teacher on well-lit image

   python tools/train_stage1_PT.py --cfg experiments/exlpose/PT_stage_config.yaml \
       DATASET.ROOT ${DATASET_ROOT} MODEL.NAME hrnet_comp

Pre-Training Stage: Training the comp. teacher on fake low-light image

   python tools/train_stage1_PT.py --cfg experiments/exlpose/PT_stage_config.yaml \
       DATASET.ROOT ${DATASET_ROOT} TRAIN.STAGE PT_LL MODEL.NAME hrnet_comp \
           MODEL.PRETRAINED_COMP ${COMP_WEIGHTS_FILE} \

KA Stage:

   python tools/train_stage2_KA.py --cfg experiments/exlpose/KA_stage_config.yaml \
       DATASET.ROOT ${DATASET_ROOT} MODEL.PRETRAINED_MAIN ${MAIN_WEIGHTS_FILE} \
           MODEL.PRETRAINED_COMP ${COMP_WEIGHTS_FILE}

Acknowledge

Our code is mainly based on DEKR.

Citation

@inproceedings{DA-LLPose,
  title={Domain-Adaptive 2D Human Pose Estimation via Dual Teachers in Extremely Low-Light Conditions},
  author={Ai, Yihao and Qi, Yifei and Wang, Bo and Chen, Yu and Wang, Xinchao and  Tan, Robby T.},
  booktitle={ECCV},
  year={2024},
}