Establish whether component removal generalizes beyond activity recognition

Determine whether removing the leading components of egocentric video embeddings improves performance on egocentric tasks other than human activity recognition.

Background

The reported component-removal results are evaluated for human activity recognition using the WEAR dataset and several pretrained video encoders. The analysis shows consistent gains across the tested encoders in this task, but the study does not evaluate other egocentric applications such as retrieval, localization, or embodied-agent tasks.

The authors explicitly identify this broader applicability as unresolved, so the finding that component removal helps activity recognition cannot be assumed to transfer to other egocentric tasks.