Dual Graph Networks for Pose Estimation in Crowded Scenes(IJCV2023)

• PyTorch(>=1.0 && <=1.13.1)
• mmcv(1.x)
• OpenCV
• visdom
• pycocotools

This code is tested under Ubuntu 16.04, CUDA 10.1, cuDNN 7.1 environment with four NVIDIA TitanXP GPUs.

Python 3.6.5 version and PyTorch 1.3.1 is used for development.

All settings about CrowdPose Datasets please refer to OPEC-Net project

Our results can be downloaded in Baidu
Extraction code: mtfc

The results.zip has two files, one is the weights we trained named best_checkpoint.pth, the other is keypoint result named crowdpose_keypoints_results.json.

verify the results
make an directory named CrowdPose_ADGN_withFlip_60epochs_bothConfidenceScoreIn_normalizedHeatmaps_batch32 under the directory checkpoints. Then, copy the weight file best_checkpoint.pth into the directory. Then one can run test script (see below) to verify our results.

training log
our training log can be seen in ./log, it is in tensorboard format. If one wants to see this result, one should install tensorboard first, then run the following command:

tensorboard --logdir CrowdPose_ADGN_withFlip_60epochs_bothConfidenceScoreIn_normalizedHeatmaps_batch32 --bind_all

e.g.

TRAIN_BATCH_SIZE=14
CONFIG_FILES=./configs/OPEC_GCN_GrowdPose_Test_FC.py
bash train.sh ${TRAIN_BATCH_SIZE} ${CONFIG_FILES} 

after training, the result of CrowdPose is save into checkpoints/name/mAP.txt
the format of results like:

epoch (without best match) (use best match) 

We also offer logs in form of tensorboard chart in log/name, it can be invoked by the command

tensorboard --logdir log/name/ --bind_all

e.g.

CHECKPOINTS_DIRS='path to your checkpoints files'
CONFIG_FILES =./configs/OPEC_GCN_GrowdPose_Test_FC.py
bash test.sh ${CHECKPOINTS_DIRS} ${CONFIG_FILES}

Note:
　　Ours (joint branch) shows the result being outputted from the joint branch of DGN.
　　Ours (after fusion of both breanches) shows the result after fusion post-process. The fusion process is detailed on the paper. Any one can code it out from the steps we offer on the paper. We will update the post-process code in the near future.

If you find our works useful in your reasearch, please consider citing:

@article{tu2023dual,
  title={Dual Graph Networks for Pose Estimation in Crowded Scenes},
  author={Jun Tu, Gangshan Wu & Limin Wang},
  journal={International Journal of Computer Vision},
  year={2023}
}