feat(reid): R6 scene-adaptive ReID accept threshold (tighten in look-alike scenes)#94
Merged
Conversation
Adds `adaptive_threshold_hi()` pure helper to reid.py that raises the recovery accept threshold toward the most-similar candidate pair (max pairwise cosine + margin), clamped to [base_hi, cap]. Wires `eff_hi` into `_maybe_reid_recover` in camera_worker so look-alike scenes demand higher confidence before rebinding — prevents ID switches without loosening the threshold in distinct scenes. Behavior is identical when <2 candidates or ReID is disabled. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Merged
TCVinNYC
added a commit
that referenced
this pull request
Jun 24, 2026
Second **pre-release** of the 2.2.0 reliability + performance roadmap, for real-hardware validation before the stable 2.2.0. Bumps `__version__` to `2.2.0-rc2`; on merge, tag `v2.2.0-rc2` triggers the pre-release installer build. **New since rc1 (all merged, CI-green + multi-agent-reviewed):** - #92 CPU subservice governor — fewer CPU spikes (phase-stagger) + cadence throttle under load - #93 preview-rate-cap (~20fps) - #90 live GPU-acceleration verdict in the Services panel - #91 OpenVINO auto-install for Intel - #94 scene-adaptive ReID threshold - #95/#96 opt-in fused TargetAssociator (off by default) - #97 batched detect_batch primitive **Validate especially:** CPU usage/spikes with all subservices running (your priority), and \`python -m autoptz --bench\` / the Services-panel verdict on Intel-Mac+AMD. To try the new tracking logic, set \`tracking.use_target_associator = true\`. Follow-ups after your validation: P4 coalescing scheduler, wire the associator ReID/pose cues + flip its default, int8 CPU quant, stable Win/Linux device binding. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The ReID recovery match used a FIXED
reid_threshold_hi(0.70). When two people in frame genuinely look alike (high inter-person appearance similarity), 0.70 can be too low and the lost target re-binds onto the wrong look-alike (an ID switch). When everyone looks distinct, 0.70 is fine.Safe scope: a pure
reid.adaptive_threshold_hi(candidates, base_hi)raises the recovery accept threshold toward the observed max inter-person cosine (clamp(max(base, max_pairwise+0.05), base, 0.95)) — demanding MORE confidence to re-bind in ambiguous scenes, but NEVER below the configured base, and only on the recovery path (target already lost). So it can only tighten matching where wrong-recovery is likely; it cannot regress vs. today. Wired into_maybe_reid_recover(both the scoring and confirmation calls);threshold_lounchanged; identical behavior with <2 candidates or ReID disabled.25 new pure tests; full suite green.
Roadmap item R6 (adaptivity). Real-scene validation (look-alike crowds vs distinct) at the final RC.
🤖 Generated with Claude Code