ZQSIAT / DOGV-ST3D-1

An implementation for "Exploiting Spatio-Temporal Representation for 3D Human Action from Depth Map Sequences"

Exploiting Spatio-Temporal Representation for 3D Human Action Recognition from Depth Map Sequences

This repository is an implementation of "Exploiting Spatio-Temporal Representation for3D Human Action Recognition from Depth Map Sequences".

Installation

1. Clone the repository https://github.com/xp-ji/DOGV-ST3D
2. Install python dependencies

• Create conda environment with dependencies: conda env create -f environment.yml
• Activate the environment: conda activate dogv

Quick Start

1. Download required dataset, for example NTU_RGB+D from the ROSE webpage. Your dataset structure should look like

ntu_rgb+d/
    depth/
    ...
    S001C003P008R001A045/
        Depth-00000001.png
        ...
        Depth-00000054.png
    ...
    depth_masked/
        ...
        S001C003P008R001A045/
            MDepth-00000001.png
            ...
            MDepth-00000054.png
        ...

2. Download the model weights from Google Drive and put the *.pth files in weights.
3. Run the command

python test_model.py --arch Resnet3d --dataset NTU_RGBD  --intype DOGV --exp X-Sub --weights "weights/NTU_RGBD_Resnet3d_X-Sub_DOGV.pth" 

The test result will be saved in logs/Test_NTU_RGBD_Resnet3d_X-Sub_DOGV.

Re-implementation

Prepare datasets

Download all needed dataset, i.e., NTU-RGB+D, PKU-MMD and UOW_combined3D

Training the model from scratch

python train_model.py --dataset NTU_RGBD --arch Resnet3d --intype RDMs --exp X-Sub

Citation

If you find this code useful for your research, please use the following BibTeX entry.

@article{ji2021exploiting,
  title={Exploiting spatio-temporal representation for 3D human action recognition from depth map sequences},
  author={Ji, Xiaopeng and Zhao, Qingsong and Cheng, Jun and Ma, Chenfei},
  journal={Knowledge-Based Systems},
  pages={107040},
  year={2021},
  publisher={Elsevier}
}