This repository contains code for LIGHTEN HOI detection pipeline, proposed in the ACM MM'20 paper: [LIGHTEN: Learning Interactions with Graph and Hierarchical TEmporal Networks for HOI in videos]().
LIGHTEN is implemented in Pytorch1.4 with CUDA-10.1 in python3.8. Other python packages can be installed using :
pip install -r requirements.txt
Download the pretrained models from the following folders : i) CAD120 checkpoints ii) V-COCO checkpoints
Set the corresponding paths to data and pre-trained models in config.py file. Hyper-paramters and model configurations can be set from this file. The directory structure after setting up looks like :
LIGHTEN-Learning-Interactions-with-Graphs-and-hierarchical-TEmporal-Networks-for-HOI/
CAD120/
checkpoints/
checkpoint_GCN_frame_detection.pth
checkpoint_GCN_segment_detection.pth
data/
training_data.p
testing_data.p
models/
V-COCO/
checkpoints/
data/
training_data.p
testing_data.p
action_index.json
Test_Faster_RCNN_R-50-PFN_2x_VCOCO_Keypoints.pkl
models/
cd CAD120/
python compute_RoI_feats.py
cd CAD120/
python train_CAD120.py
python test_CAD120.py
cd V-COCO/
python compute_RoI_feats.py
cd V-COCO/
python train_VCOCO.py
python eval_VCOCO.py
python test_VCOCO.py