Skip to content

Latest commit

 

History

History
54 lines (43 loc) · 2.94 KB

README.md

File metadata and controls

54 lines (43 loc) · 2.94 KB
Raw

NTU RGB+D Dataset Action Recognition with GNNs and CNNs

graph neural networks(tensorflow) and spatial convolutional neural networks(pytorch) for action recognition from NTU RGB+D skeletion dataset. The spatial convolutional neural networks use a novel virtual radar method to convert graph data in the NTU dataset into spectrograms (images).

Directory Structure

  • data_gen/: These scripts are used to convert NTU dataset from their native format to numpy tensors.

    • data_gen/gen_joint_data.py: Converts NTU data from skeleton format to joint data and saves it in numpy file

    • data_gen/gen_bone_data.py: Computes bone data from joint data

    • data_gen/gen_motion_data.py: Computes motion data from joint and bone data

    • data_gen/gen_tfrecord_data.py: Converts prepossessed numpy dataset to tensorflow's tfrecord fromat. Needed to train with GNNs in this repo

  • graph/: Methods to define graph adjacency or incidence matrix for graph neural networks. Incidence'

  • layers/virtual_radar.py: Custom pytorch layer to compute spectrograms from temporal graph data by simulating a virtual radar to capture the graph's motion

  • models/: Neural network definitions. All Graph Neural Networks are defined in tensorflow. Vanilla Spatial Convolutional Networks are defined in pytorch.'

  • main_gnn.py: Script to train on NTU dataset with graph neural networks

  • main_spectrogram.py: Script to train on NTU dataset with spatial convolutional neural networks. Since the NTU data is natively represented as graphs, it is converted to spectrogram(images) with the custom VirtualRadar layer.

  • virtual_radar_example.ipynb: Jupyter notebook with an example of VirtualRadar layer applied to skeleton data from Microsoft Kinect Azure camera.

NTU RGB+D dataset

The NTU RGB+D dataset can be downloaded from here. We'll only need the Skeleton data (~ 5.8G).

The dataset is very noisey, consider adding methods clean it. Checkout VideoPose3D, might be able to use video data from NTU dataset to generate skeleton data and combine it with skeleton data from NTU dataset which uses depth data to get skeletons.

References