This repository will contain the code of the paper "Glimpse Clouds: Human Activity Recognition from Unstructured Feature Points" by F. Baradel, C. Wolf, J. Mille and G.W. Taylor presented at CVPR 2018.
Check out the project page for more information.
You should download the NTU-DATASET and install the lintel module for efficiently decoding videos on the fly.
Check the dataprocessing README for more information about how to rescale the videos.
Run the following python command to make sure you are able to iterate over the videos.
python ./src/data/ntu/test.py
The training is split into two steps:
- We train the 3D-CNN using global average pooling
- We freeze the backbone and train the Glimpse Clouds This can be done by running the following command:
./script/train_and_evaluate.sh <DIR-CKPT> <LOCATION-NTU> where <DIR-CKPT> is where you want to store your files (log, checkpoints) and <LOCATION-NTU> is the location of the NTU dataset on your cluster.
Below is the generic python command for running the code
python src/main.py \
--dir <DIR-CKPT> \
--root <DATA> \
-b <BATCH-SIZE> \
-t <NB-TIMESTEPS> \
-lr <LEARNING-RATE> \
--epochs <EPOCHS> \
-j <NB-WORKERS> \You need to adding either --global-model or --glimpse-clouds for training respectively the backbone-CNN or the glimpse clouds.
Pose predictions can be added by adding the option --pose-predictions.
Evaluation can be done by adding the option -e.
- python 3 and higher
- pytorch 0.4
- lintel
If our paper and/or this code is useful for your research, please consider citing us:
@InProceedings{Baradel_2018_CVPR,
author = {Baradel, Fabien and Wolf, Christian and Mille, Julien and Taylor, Graham W.},
title = {Glimpse Clouds: Human Activity Recognition From Unstructured Feature Points},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2018}
}
MIT License