OPEC-GCN: Occluded Pose Estimation and Correction using Graph Convolutional Neural Networks

In our work, we mainly use three dataset to evaluate our considerablely results.
you can download the dataset from below link.
CrowdPose
OCHuman
MSCOCO

Firstly you should download the dataset, and then your project folder looks like follow structure. All of img_dir you can modify in config

--Crowdpose/images/...
--train2017/...
--OPEC-GCN/...

At the same time, you need download the weights of sppe and yolov3 because our OPEC-GCN depends on Alphapose as base module. So please download the models manually: duc_se.pth (2018/08/30) (Google Drive | Baidu pan), yolov3-spp.weights(Google Drive | Baidu pan). Place them into ./weights/sppe and ./weights/yolo respectively.

Considering convenience, I already processed datasets for you so that you can easily train your opec-gcn model. download the json file manually: train_process_datasets, test_process_datasets. Place them into ./train_process_datasets and ./test_process_datasets respectively.

You can easily start to train CrowdPose dataset using following code.

CUDA_VISIBLE_DEVICES=0 python ./tools/train_alpha_pose_gcn.py --indir ../crowdpose/images/ --nEpochs 25 --trainBatch 20 --validBatch 60 --LR 1e-3 --dataset 'coco' --config ./configs/OPEC_GCN_CrowdPose_Test.py

Results on CrowdPose-test datasets:

Results on OCHuman datasets:

Results on OCPose datasets:

Left is current state of the art method, other side is our method.