acceptodds
Desk rejected at ICLR 2027

Same Object, or Its Twin? Diagnosing Instance Association Across Scene Revisits

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Persistent object maps must maintain physical identity across changes in position and viewpoint. An incorrect association can reflect absent observational cues, difficulty reading them from features, or limitations in how matching uses them. We test these explanations through simulated paired revisits of otherwise identical chairs with artificial persistent color markers. Representations share oracle masks, views and the full candidate map. In the primary strong-color swap condition with correlated cameras, an RGB descriptor recovers most identities. A development-fitted linear probe decodes marker colors from MASA-R50 features, yet maximum view-pair cosine matching frequently selects the exchanged partner. Holding view features and candidates fixed, normalized-mean aggregation increases correct associations from 2 to 10 of 36 queries. Penalizing differences in the probe's continuous output raises this count to 15. Aggregation accounts for most of both this recovery and the accuracy loss on other map objects, while readout gains vary across representations and camera conditions. This partial recovery shows that evaluating readable identity cues requires testing how fixed features are aggregated and compared under full-map candidate competition.

Primary area:
applications to robotics, autonomy, planning
Keywords:
object data association, persistent object mapping, object identity, identifiability, embodied perception, 3D scene understanding, selective prediction, semantic SLAM

Related papers

Discussion (0)

Sign in to comment.