This is the code repository of the final project for the 19Fall Computer Vision Course (EN.601.661) at JHU. The team members are Heather Han, Zili Huang, Yingda Xia and Yi Zhang. The code is based on MMAction.
The master branch is for RGB modality. Checkout the optical_flow
and rgb+kp
branch for optical flow modality and human 2D keypoint modality.
Please refer to INSTALL.md for installation.
We use a subset of NTU RGB+D dataset. We provide a script to process the dataset and generate necessary files for training and testing.
bash prepare_nturgbd.sh
We provide pretrained models for testing. Download them to modelzoo/
,
bash download_models.sh
Test models on the testset,
bash test_rgb.sh
To ensemble the results of different modalities, we use late fusion which averages the logits from the output of different models.
The output results in our experiment are saved in results/
. Ensembling using the following script,
python ensemble.py
We provide a script for training RGB network.
bash train_rgb.sh