Accepted by 17th European Conference on Computer Vision (ECCV2022)
Pytorch Implementation of Video Activity Localisation with Uncertainties in Temporal Boundary
- Clone this repo:
git clone https://github.com/Raymond-sci/EMB.git
- Download the pre-trained video features from here and word embeddings from here, then put them in
data/features
. - Set up experimental environments using
environment.yml
- Download our trained models from here and put them in
sessions/
, create the session folder if not existed. - Run
python main.py --task charades --mode test --model_name 20220715-202236
to evaluate a model. By default, the determined boundaries predicted by the model will be tested, use--elastic
if want to test the elastic boundaries. - Run
python main.py --task charades --mode train
to train on one of{charades, tacos, activitynet}
benchmarks.- by default, no intermediate files (log, checkpoints and etc) will be stored on disk during training, use
--deploy
if needed - Use the option
--model_dir
to specify where the files generated by experiment sessions should be stored. The default path issessions/
. - Use the option
--model_name
to specify the session name. The timestamp will be used as the default session name. - More options can be found in
main.py
- by default, no intermediate files (log, checkpoints and etc) will be stored on disk during training, use
This project is licensed under the MIT License. See LICENSE for more information
Please cite our paper if you found this project helpful:
@InProceedings{huang2022emb,
title = {Video Activity Localisation with Uncertainties in Temporal Boundary},
author = {Jiabo Huang, Hailin Jin, Shaogang Gong, Yang Liu},
booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
year = {2022},
}
This implementation is heavily based on the excellent work VSLNet carried out with the paper Span-based Localizing Network for Natural Language Video Localization.