Model

ReID + Gait

Re-identify subjects across cameras and time to build long-term dossiers.

shipped Re-identification IdentityEmbeddings

Overview

Appearance embeddings, soft biometrics (colour, height, accessories) and z-scored, winsorized gait descriptors are fused into a subject signature. Cosine matching links sightings across cameras and days, powering repeat-visitor detection, movement trails, relationship graphs and the identity dossier.

RoleRe-identification
Expected latency~10–25 ms / embedding (GPU)
Hardware requirementsNVIDIA GPU recommended; embeddings serialised to avoid contention
Memory requirements~1 GB VRAM

Inputs & Outputs

Inputs
  • Person crop
  • Track id
  • Gait sequence
Outputs
  • Appearance embedding
  • Subject uid + match score
  • Soft-biometric attributes

Advantages & Limitations

Advantages

  • Cross-camera + cross-day identity
  • Gait adds a hard-to-spoof cue
  • Feeds dossiers and relationship graphs

Limitations

  • Appearance drifts with clothing / lighting
  • Gait needs a clean walking sequence
  • Probabilistic, surfaced with confidence, not certainty

Example outputs

Configuration

Keys in config/default.yaml that govern this model.

KeyDefault
roster.persisttrue
roster.persist_threshold0.74
gait.enabledtrue

Benchmarks

Placeholder

Benchmark figures below are illustrative and awaiting a standardised harness. Treat them as placeholders.

BenchmarkDatasetMetricValueHardware
Rank-1internalaccuracyplaceholder-