This code will be published on May 1, 2021
Paper : DarkLight Networks for Action Recognition in the Dark | IEEE Conference Publication | IEEE Xplore
DarkLight Networks,a high-efficiency 3D Convolution network for low light illumination action recognition.
This repository is modeling on the foundation of LateTemporalModeling3DCNN architectures and taking advantage of self-attention to build the classification network.
The code runs on Python 3.8 but in fact it is not a big deal to try to construct a low version Python. You can create a conda environment with all the dependecies by running
conda env create -f requirements.yml -n <env_name>
Note: this project needs the CUDA 11.1
In order to improve the dataset's load speed an facilitate pretreatment. The video should be pre-cut into frames and save in \datasets folder and the list of training and validation samples should be created as txt files in \datasets\settings folder.
For example, using ARID as the data. the frames is saved in \datasets\ARID_frames. And then using the different name of classifications to name the folder.
The format of the frames like as "img_%05d.jpg"
training with attention
python darklight.py --split=1 --batch-size=8 --gamma=1.8 --both-flow = True
training without attention
python darklight.py --split=1 --batch-size=8 --gamma=1.8 --both-flow = True --no_attention
For multi-gpu training, comment the two lines below os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID" os.environ["CUDA_VISIBLE_DEVICES"]="0"
To continue the training from the best model, add -c. To evaluate the single clip single crop performance of best model, add -e
python spatial_demo_bert.py --split=1
R2+1D-IG65: IG-65M Pytorch
self-attention-cv:self-attention-cv Pytorch
LateTemporalModeling3DCNN:2_Plus_1D_BERT