Official implementation in python. https://arxiv.org/pdf/2003.13158.pdf
If you use the code, please cite
@inproceedings{kukleva2020lirec,
title={Learning Interactions and Relationships between Movie Characters},
author={Kukleva, Anna and Tapaswi, Makarand and Laptev, Ivan},
booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR'20)},
year={2020}
}
To run the code first download the data: https://www.rocq.inria.fr/cluster-willow/mtapaswi/downloads/anna_cvpr2020/
Note that it's about 80GB. Put all the downloads in the same folder data_root
parameter (see below) with the subfolders:
- models_release/
- features/
- dialogs/
- frame2time/
- ftracks/
- ftrack_ids/
- intersections/
- others/
- Set all the paths:
LIReC/utils/arg_pars.py
--project_root: path to the LIReC project
--data_root: path to the folder with data
--store_root: just in case if you want to try training to store models (optional)
- Install environment
conda create --name lirec --file requirements.txt
conda actiavte lirec
- To resume models from checkpoints:
modality check (model there is only for all three modalities)
python resume/modalities.py
multi-task learning for interactions and relationships
python resume/int_rels.py
interactions and movie characters detection
python resume/int_ch.py
interactions, relationships and movie characters detection
python resume/int_rel_ch.py
Each file has option sanity_check
. If it is set to True, you can quickly check if nothing breaks with the data paths and models.
If it is set to False, test will be made on the entire dataset.
-
Movie character detection can be evaluated with model trained on ground truth or weakly trained model. Set to corresponding value
tr_correct
in 'resume/int_ch.py' or 'resume/int_rel_ch.py'. -
No specific code for training these models, sorry. But you can find trainig function, all the losses and other details in the code. If any questions, just drop me an email.