README.md

Code for the paper 'Modality Distillation with Multiple Stream Networks for Action Recognition'.

arXiv

ECCV 2018

ERRATA (ECCV paper, Figure 2): first convolutional block of each ResNet unit is 1x1, not 3x3. The code is correct.

If you use this code as part of any published research, please acknowledge the following paper:

@InProceedings{Garcia_2018_ECCV,
author = {Garcia, Nuno C. and Morerio, Pietro and Murino, Vittorio},
title = {Modality Distillation with Multiple Stream Networks for Action Recognition},
booktitle = {The European Conference on Computer Vision (ECCV)},
month = {September},
year = {2018}
}

Data and Imagenet checkpoint directories are defined in utils.py

Arguments are described in utils.py - get_arguments().

To read logs

cat ./log/uwa3dii/s1_train_depth_01012018_010101__dset_uwa3dii_eval_mode_cross_view/log.txt

Train Step 1 models, e.g.

depth: python s1_train_stream.py --dset=ntu --modality=depth --eval=cross_subj

RGB: python s1_train_stream.py --dset=nwucla --modality=rgb

Train Step 2 model, e.g.

Define the right path for depth and rgb checkpoints in s2_twostream_depth_rgb.py python s2_twostream_depth_rgb.py --interaction --dset=ntu-mini

Evaluate twostream model, rgb and depth e.g.

python s2_twostream_depth_rgb --dset=uwa3dii --just_eval --ckpt=./step2_checkpoint_dir_model.ckpt

Train Hallucination model, e.g.

python s3_distillation.py --dset=nwucla --ckpt=./step2_checkpoint_dir_model.ckpt

Train Step 4 model, e.g.

Define the right path for depth (checkpoint from step 3) and rgb (checkpoint from step 2) ckpts in s4_depth_hall.py python s4_depth_hall.py --dset=uwa3dii

Evaluate Step 4 model (rgb and hall), e.g.

python s4_depth_hall.py --dset=uwa3dii --just_eval --ckpt=./step4_checkpoint_dir_model.ckpt