Whareformer: A New Approach for Online Object Tracking in Long Egocentric Videos
Whareformer is an online learning object tracking method designed for long egocentric videos, with its paper accepted by ECCV 2026. It addresses the challenge of tracking objects that disappear and reappear over extended periods from a first-person perspective.
Key Details and Architecture
The model aims to solve the Object Search and Tracking task in long egocentric videos (OSNOM). Its core architecture trains an embedding network that combines object appearance with 3D position differences, utilizing a transformer encoder for contrastive allocation. For memory, the model introduces explicit new track tokens and uses the online DenStream algorithm to efficiently maintain persistent and transient appearance micro-clusters. This online assignment-based memory mechanism never deletes trajectories, explicitly retaining out-of-sight objects to enable long-term reuse and tracking.
Performance and Impact
Although trained on only 50 videos from the EPIC-KITCHENS dataset, Whareformer performed excellently across over 100 long video tests, surpassing other alternatives and demonstrating high data efficiency and robustness in long-video tracking.
2026-07-10 ~ 2026-07-10 · 6 related posts
- [source] Whareformer: Online Object Tracking for Long Egocentric Videos — dimadamen · 2026-07-10
- Whareformer: A New Architecture for Object Tracking in Long Egocentric Videos — dimadamen · 2026-07-10
- Whareformer: A Novel Approach for Long Video Object Tracking — dimadamen · 2026-07-10
- [source] Whareformer: A Novel Approach for Long Video Object Tracking — dimadamen · 2026-07-10