This is a visualize version of OPEC-Net.

Peeking into occluded joints: A novel framework for crowd pose estimation(ECCV2020)

• Due to the memory of GPU, I have only tested the OPEC_GCN_CrowdPose_Test_FC.py model.
• You could get better result if you choose other config model to train.
• The wild img / video result and demo script will be updated soon !!

Make sure you have installed the python UV first

  uv.sync

This code is tested under win 11 environment with single NVIDIA 4060 GPU.

Python 3.9 version is used for development.

My trained checkpoints is here

Alphapose+

SPPE

yolov3

Original author provide the download link of OCPose dataset.

Google Drive

• Images
• Annotations

Baidu Drive

• Image code:euq6
• Annotations code:3xgr

• COCO2017
• CrowdPose

pls, Download annotations processed by sampling rules according to our paper

• train_process_datasets
• test_process_datasets

The original and optimized json file for visualization.

• ori_json

• pred_json

Your project structure should be like this:

${Project_dir}
└───data
│   └───coco
│   │      train2017
│   │      └─── xx.jpg
│   └───crowdpose
│          └─── xx.jpg
│   └───images
│          └─── xx.jpg
└───train_process_datasets
│      │--  train_crowd_train.json
│      └─── coco_pose_45.json
└───test_process_datasets
│      │--  crowdpose_test.json
│      │--  test_compute_map++.json
│      └─── pred_test_best_match.json
└───vis_process_datasets
│      │--  pred.json
│      └─── ori.json
└───weights
       │-- sppe
       │     sppe weights
       |-- ssd
       └───yolo 
             yolow eights

python vis_crowd.py

The result imgs will be saved in vis_crowd folder.

Refer to the original repo.

• OPEC-Net
• AlphaPose

[email protected]