This repository contains the implementation for ICLR2024 paper LLCP: Learning Latent Causal Processes for Reasoning-based Video Question Answer pdf

LLCP is a causal framework designed to enhance video reasoning by focusing on the spatial-temporal dynamics of objects within events, without the need for extensive data annotations. By employing self-supervised learning and leveraging the modularity of causal mechanisms, LLCP learns multivariate generative model for spatial-temporal dynamics and thus enables effective accident attribution and counterfactual prediction of Reasoning-based VideoQA.

First, please install the recent version of Pytorch and Torchvision as pip install torch torchvision. Then, you can install other package by running pip install -r requirements.txt

We provide the processed features used in our experiments. Please download the data and model in this link1 and this link2. Then please decompress the floders as ./data/ and ./results/ and replace the original floders as the downloaded ones.

The directory structure should look like

LLCP_VQA/
|–– config.py
|–– configs/
|–– data/
|   |–– object_test_feat/
|   |–– object_train_feat/
|   |–– appearance_feat_rn50.h5
|   |–– test_questions.pt
|   |–– train_questions.pt
|   |–– video_noaccident_train.txt
|–– DataLoader.py
|–– models_cvae.py
|–– requirements.txt
|–– results/
|   |–– .../model_cvae49.pt
|–– README.md
|–– train.py
|–– validate.py

To train the cvae model, you can run this command:

python train.py --cfg configs/sutd-traffic_transition.yml

To evalaute the trained model, please refer to:

python validate.py --cfg configs/sutd-traffic_transition.yml

See LLCP-Simulation.

If you find our work useful in your research, please consider citing:

@inproceedings{chen2024llcp,
  title={LLCP: Learning Latent Causal Processes for Reasoning-based Video Question Answer},
  author={Chen, Guangyi and Li, Yuke and Liu, Xiao and Li, Zijian and Al Surad, Eman and Wei, Donglai and Zhang, Kun}
  booktitle={ICLR},
  year={2024}
}

Our implementation is mainly based on the SUTD-TrafficQA and Tem-Adapter, we thank the authors to release their codes.