简体中文 | English

AGCN

---

Contents

• Introduction
• Data
• Train
• Test
• Inference
• Reference

Introduction

We implemented Adaptive Graph Convolution Network to improve the accuracy of ST-GCN.

Data

Please refer to FSD-10 data download and preparation doc FSD

Please refer to NTU-RGBD data download and preparation doc NTU-RGBD

Train

Train on FSD

• Train AGCN on FSD scripts:

python3.7 main.py -c configs/recognition/agcn/agcn_fsd.yaml

• Turn off valid when training, as validation dataset is not available for the competition.

Train on NTU-RGBD

• Train AGCN on NTU-RGBD scripts:

python3.7 -B -m paddle.distributed.launch --gpus="0,1,2,3"  --log_dir=log_agcn  main.py  --validate -c configs/recognition/agcn/agcn_ntucs.yaml

• config file agcn_ntucs.yaml corresponding to the config of AGCN on NTU-RGB+D dataset with cross-subject splits.

Test

Test onf FSD

• Test scripts：

python3.7 main.py --test -c configs/recognition/agcn/agcn_fsd.yaml  -w output/AGCN/AGCN_epoch_00100.pdparams

• Specify the config file with -c, specify the weight path with -w.

• Evaluation results will be saved in submission.csv file, final score can be obtained in competition website.

Accuracy on FSD dataset:

Test_Data	Top-1	checkpoints
Test_A	62.29	AGCN_fsd.pdparams

Test on NTU-RGB+D

• Test scripts：

python3.7 main.py --test -c configs/recognition/agcn/agcn_ntucs.yaml -w output/AGCN/AGCN_best.pdparams

• Specify the config file with -c, specify the weight path with -w.

Accuracy on NTU-RGB+D dataset:

split	Top-1	checkpoints
cross-subject	83.27	AGCN_ntucs.pdparams

Inference

export inference model

To get model architecture file AGCN.pdmodel and parameters file AGCN.pdiparams, use:

python3.7 tools/export_model.py -c configs/recognition/agcn/agcn_fsd.yaml \
                                -p data/AGCN_fsd.pdparams \
                                -o inference/AGCN

• Args usage please refer to Model Inference.

infer

python3.7 tools/predict.py --input_file data/fsd10/example_skeleton.npy \
                           --config configs/recognition/agcn/agcn_fsd.yaml \
                           --model_file inference/AGCN/AGCN.pdmodel \
                           --params_file inference/AGCN/AGCN.pdiparams \
                           --use_gpu=True \
                           --use_tensorrt=False

example of logs:

Current video file: data/fsd10/example_skeleton.npy
        top-1 class: 27
        top-1 score: 0.8965644240379333

Reference

• Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition, Sijie Yan, Yuanjun Xiong, Dahua Lin

• Two-Stream Adaptive Graph Convolutional Networks for Skeleton-Based Action Recognition, Lei Shi, Yifan Zhang, Jian Cheng, Hanqing Lu

• Skeleton-Based Action Recognition with Multi-Stream Adaptive Graph Convolutional Networks, Lei Shi, Yifan Zhang, Jian Cheng, Hanqing Lu

• Many thanks to li7819559 and ZhaoJingjing713 for contributing the code.