This project adapts MeteorNet and Graph-RNN to human action recognition. The structure of the codes are shown below:
├─action_cls
│ └─log
│ ├─test
│ └─train
├─chain_interp_flow_preprocess
├─Depth
├─tf_ops
│ ├─3d_interpolation
│ ├─grouping
│ └─sampling
└─utils
Models are put in folder action_cls. The core part of models, such as the Meteor module and Graph-RNN module are put in utils. The dataset should be put in Depth. Chain_interp_flow_preprocess is the grouping code for MeteroNet, and tf_ops are codes for TF operators.
These codes are running on:
Python 3.5
Tensorflow-gpu 1.9.0
CUDA 11.6.1
OpenCV
The TF operators are included under tf_ops, you need to compile them first by make under each ops subfolder. Make sure to change the CUDA_HOME as the same as yours. Change the arch version according to your CUDA. For example, our cuda Tesla V100, the corresponding arch version is sm_70. You can check CUDA Compute Capability to find the version that suits your GPU.
Download raw MSRAction3D dataset from here (~62MB). Extract the .zip file to get Depth.rar file and extract it in Depth directory. Then in this directory, run the following command to preprocess the data. --num_cpu flag is used to specify the number of CPUs to use during parallel processing.
python preprocess_file.py --input_dir /path/to/Depth --output_dir processed_data --num_cpu 11
The script for training and testing the model is command_train.sh. To train, use the following command.
sh command_train.sh
One may change the flags such as num_frame, num_point etc for different architecture specs.
If you want to train MeteorNet, change model to model_cls_direct;
If you want to train Graph-RNN, change model to model_cls_graphrnn_v2.