Evidence-based re-identification: Leveraging appearance cues to correct tracking mistakes rather than trying to prevent them

Evidence-based re-identification keeps identity tracking reliable for minutes-long recordings of mice in enriched environments, with ByteTrack+re-identification outperforming ByteTrack alone and Tailtag.

• Why it matters: Frame-by-frame multi-animal tracking can lock in even sporadic identity swaps, making error-free results progressively less likely as recordings lengthen and cohorts grow.
• What they did: The method builds appearance representations per subject and continuously monitors their evolution, correcting when two representations drift toward each other in a sustained way. It was tested on two distinct minutes-long mouse benchmarks using DeepLabCut, SLEAP, ByteTrack, and BoT-SORT-ReID.
• The result: Motion-only trackers and BoT-SORT-ReID accumulated identity errors within seconds, while ByteTrack with evidence-based re-identification remained reliable and was positively correlated with appearance-extractor accuracy only under this re-identification scheme.

AI summary of the abstract · may contain errors · not medical advice · read the paper.