EPIC-KITCHENS-100 Challenge
Two-time consecutive winner of the Unsupervised Domain Adaptation Challenge
[ACHIEVEMENT] 2021–2022
EPIC-KITCHENS-100 Challenge 🏆
Two-Time Consecutive Winner of the Unsupervised Domain Adaptation Challenge at CVPR — competing against leading research institutions worldwide.
ACHIEVEMENT_SUMMARY
3rd place in 2021 and Top 3 in 2022 at the EPIC-KITCHENS Unsupervised Domain Adaptation Challenge.
2022 — CVPR Workshop
TOP 3 WINNEROur "Multi-Source Multi-Target Unsupervised Domain Adaptation" approach secured a position among the top 3 winner teams at the EPIC-KITCHENS UDA Challenge presented at CVPR 2022.
Challenge Context
The challenge focused on multi-source multi-target domain adaptation, testing how models generalize across different kitchen environments and temporal shifts in egocentric action recognition.
Key Innovation: We developed a novel approach to leverage multiple source domains while adapting to multiple unlabeled target domains simultaneously, achieving robust performance across all evaluation metrics.
2021 — CVPR Workshop
3RD PLACETogether with Chiara Plizzari, we achieved 3rd place in the third edition of the challenge, presented at the CVPR 2021 Workshop on Egocentric Perception, Interaction and Computing.
Technical Approach
We re-purposed the Relative Norm Alignment (RNA) loss, a multi-modal loss recently proposed for Domain Generalization, to operate between different backbone architectures in order to enhance their collaboration.
Results: Top performance across all evaluation categories: verb, noun, and action recognition.
$ info --dataset EPIC-KITCHENS-100
About the Dataset
EPIC-KITCHENS-100 is the largest-scale egocentric dataset, collected by 32 participants in their native kitchen environments, densely annotated with actions and object interactions.
Unsupervised Domain Adaptation Challenge
The challenge tests how models cope with temporal domain shift in action recognition. Videos recorded in 2018 (EPIC-KITCHENS-55) constitute the labeled source domain, while videos recorded two years later (EPIC-KITCHENS-100 extension) constitute the unlabeled target domain.
Objective: Assign accurate (verb, noun) labels to trimmed segments following the Unsupervised Domain Adaptation paradigm.