Code for the paper MMTM: Multimodal Transfer Module for CNN Fusion. This is a reimplementation of the original MMTM code to reproduce the results on NTU RGB+D dataset in Table 5 of the paper.

If you use this code, please cite the paper:

@inproceedings{vaezi20mmtm,
 author = {Vaezi Joze, Hamid Reza and Shaban, Amirreza and Iuzzolino, Michael L. and Koishida, Kazuhito},
 booktitle = {Conference on Computer Vision and Pattern Recognition ({CVPR})},
 title = {MMTM: Multimodal Transfer Module for CNN Fusion},
 year = {2020}
}

This code has been tested on Ubuntu 16.04 with Python 3.8.3 and PyTorch 1.5.0.

• Install Pytorch.
• Install tqdm by running pip install tqdm.
• Install opencv by running pip install opencv-python.
• Install matplotlib by running pip install matplotlib.
• Install sklearn by running pip install sklearn.

• Clone this repository along with MFAS submodule by running git clone --recurse-submodules https://github.com/haamoon/mmtm.git
• Download and uncompress the checkpoints and place them in 'ROOT/checkpoint' dicrectory.
• Download NTU RGB+D dataset.
• Copy all skeleton files to ROOT/NUT/nturgbd_skeletons/ directory.
• Change all video clips resolution to 256x256 30fps and copy them to ROOT/NTU/nturgbd_rgb/avi_256x256_30/ directory.

• Run python main_mmtm_ntu.py --datadir ROOT/NTU --checkpointdir ROOT/checkpoints --test_cp fusion_mmtm_epoch_8_val_loss_0.1873.checkpoint --no_bad_skel.
• Reduce the batch size if run out of memeory e.g. --batchsize 1.
• Add --use_dataparallel to use multiple GPUs.

• Run python main_mmtm_ntu.py --datadir ROOT/NTU --checkpointdir ROOT/checkpoints --train --ske_cp skeleton_32frames_85.24.checkpoint --rgb_cp rgb_8frames_83.91.checkpoint.
• We have trained the model with --batchsize 20 --use_dataparallel options on 4 GPUs.