Skip to content

Synchronization is All You Need: Exocentric-to-Egocentric Transfer for Temporal Action Segmentation with Unlabeled Synchronized Video Pairs [ECCV, 2024]

License

Notifications You must be signed in to change notification settings

fpv-iplab/synchronization-is-all-you-need

Repository files navigation

Synchronization is All You Need:

Exocentric-to-Egocentric Transfer for Temporal Action Segmentation with Unlabeled Synchronized Video Pairs

This repository hosts the code related to the following paper: Camillo Quattrocchi, Antonino Furnari, Daniele Di Mauro, Mario Valerio Giuffrida, Giovanni Maria Farinella: "Synchronization is All You Need: Exocentric-to-Egocentric Transfer for Temporal Action Segmentation with Unlabeled Synchronized Video Pairs" (ECCV 2024). Download

The code in this repository is based on this repository: Repository

If you use the code hosted in this repository, please cite the following paper:

@article{quattrocchi2023synchronization,
  title={Synchronization is All You Need: Exocentric-to-Egocentric Transfer for Temporal Action Segmentation with Unlabeled Synchronized Video Pairs},
  author={Quattrocchi, Camillo and Furnari, Antonino and Di Mauro, Daniele and Giuffrida, Mario Valerio and Farinella, Giovanni Maria},
  journal={arXiv preprint arXiv:2312.02638},
  year={2023}
}

Contents

Problem Definition

Proposed Method

Data

Per-frame features are required as input. The features used in this work were extracted using DINOv2 (dinov2_vitl14, 1024-D). Link to the DINOv2 repository: DINOv2

An example of code used to extract features is shown here: DINOv2_feature_extractor.py

Run data/data_stat.py to generate data statistics for each video.

Training

To evaluate the trained models, first replace the model and feature paths within the main_{oracle/transfrormer/distillation}.py, transformer_and_distillation.py e dataset_{oracle/distillation}.py codes.

To train your model, run:

python main_{oracle/transfrormer/distillation}.py --action train --feature_path lmdb_path --split train

Or:

python transformer_and_distillation.py --action train --feature_path lmdb_path --split train

Evaluate

To evaluate the trained models, first replace the model and feature paths within the main_{oracle/transfrormer/distillation}.py, transformer_and_distillation.py e dataset_{oracle/distillation}.py codes.

To evaluate the trained models:

python main_{oracle/transfrormer/distillation}.py --action predict --feature_path lmdb_path --test_aug 0

Or:

python transformer_and_distillation.py --action predict --feature_path lmdb_path --test_aug 0

Ackowledgements

This research has been supported by the project Future Artificial Intelligence Research (FAIR) – PNRR MUR Cod. PE0000013 - CUP: E63C22001940006

About

Synchronization is All You Need: Exocentric-to-Egocentric Transfer for Temporal Action Segmentation with Unlabeled Synchronized Video Pairs [ECCV, 2024]

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages