PoliTO-IIT Submission to the EPIC-KITCHENS-100 Unsupervised Domain Adaptation Challenge for Action Recognition
This work addresses domain shift in action recognition for video datasets, but it is incremental as it builds on existing methods like RNA and TA3N.
The paper tackled the problem of unsupervised domain adaptation for action recognition in the EPIC-Kitchens-100 challenge by extending a domain generalization technique with new multi-stream consistency losses, achieving first place for 'verb' and third place for 'noun' and 'action' on the leaderboard.
In this report, we describe the technical details of our submission to the EPIC-Kitchens-100 Unsupervised Domain Adaptation (UDA) Challenge in Action Recognition. To tackle the domain-shift which exists under the UDA setting, we first exploited a recent Domain Generalization (DG) technique, called Relative Norm Alignment (RNA). It consists in designing a model able to generalize well to any unseen domain, regardless of the possibility to access target data at training time. Then, in a second phase, we extended the approach to work on unlabelled target data, allowing the model to adapt to the target distribution in an unsupervised fashion. For this purpose, we included in our framework existing UDA algorithms, such as Temporal Attentive Adversarial Adaptation Network (TA3N), jointly with new multi-stream consistency losses, namely Temporal Hard Norm Alignment (T-HNA) and Min-Entropy Consistency (MEC). Our submission (entry 'plnet') is visible on the leaderboard and it achieved the 1st position for 'verb', and the 3rd position for both 'noun' and 'action'.