Users praised the Whareformer paper's fantastic collaboration between Bristol CS and Naver Labs Europe for advancing 3D reasoning to track dynamic objects in long egocentric videos.
Based on 1 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.
Whareformer @eccvconf #ECCV2026 fantastic collab bw @bristolcs @naverlabseurope led by @JacobChalkie w/ @sinha_saptarshi Yannis Kalantidis & @dlarlus Paper, Code &models to reproduce public https://jacobchalk.github.io/Whareformer/ https://arxiv.org/abs/2607.08537 N/N https://x.com/dimadamen/status/2075531087711781301/video/1
NEW Whareformer: Learning to Track What is Where in Long Egocentric Videos @eccvconf #ECCV2026 paper The first online learning approach to track dynamic objects in ego video reasoning on appearance &3D location w track memory & explicit new track token. https://jacobchalk.github.io/Whareformer/ 🧵 https://x.com/dimadamen/status/2075527352608592184/video/1
Whareformer (What&Where former) tackles the OSNOM task - Out of Sight objects remain explicitly in memory so you know where objects are at all times. In natural ego videos, tracking moving objects offers a significant challenge, compared to tracking static parts of the scene. 2/N https://x.com/dimadamen/status/2075528377633513535/video/1
We train: * embedding network g(.) to combine appearance &location diff * NT token to learn to initialise new tracks * transformer encoder to contrast assignments Memory is updated online based on assignments. No tracks are removed - in ego you reuse objects after LONG gaps 3/N https://x.com/dimadamen/status/2075528490238005739/photo/1
We utilise the online DenStream algorithm to maintain persistent (recurrent viewpoints) and transient (new viewpoints) appearance micro-clusters within each track efficiently. Unlike applications in autonomous or multi-person tracking, tracks can be revisited after 30+mins 4/N https://x.com/dimadamen/status/2075528634442396132/video/1
We utilise the online DenStream algorithm to maintain persistent (recurrent viewpoints) and transient (new viewpoints) appearance micro-clusters within each track efficiently. Unlike applications in autonomous or multi-person tracking, tracks can be revisited after 30+mins 4/N https://x.com/dimadamen/status/2075528634442396132/video/1
Whareformer is only trained on 50 videos from EPIC-KITCHENS and evaluated thoroughly on 100+ long videos from EPIC-KITCHENS, HD-EPIC and IT3Ego consistently outperforming alternatives. Watch quali: https://youtu.be/tfJ1yhIE8cE 5/N
Users praised the Whareformer paper's fantastic collaboration between Bristol CS and Naver Labs Europe for advancing 3D reasoning to track dynamic objects in long egocentric videos.
Based on 1 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.
Whareformer is only trained on 50 videos from EPIC-KITCHENS and evaluated thoroughly on 100+ long videos from EPIC-KITCHENS, HD-EPIC and IT3Ego consistently outperforming alternatives. Watch quali: https://youtu.be/tfJ1yhIE8cE 5/N