Repository navigation
fix(tracking): require unique instance support for pose detections - #132
Merged
Merged
Conversation
austinjiann
changed the base branch from
codex/all-people-identity-guard
to
main
November 21, 2026 04:18
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 city tracking baseline treated street lamps as people, split the main walker across IDs, and let two poses claim the same person mask. Unmatched MultiHMR poses previously received rectangular fallback masks even when Mask R-CNN was enabled.
Match poses to independent person instances one-to-one before temporal association. Maximize supported pairs, then overlap, at the unchanged strict IoU > 0.2 threshold. Unassigned poses cannot create tracks in the default mode; explicit no-mask mode retains its behavior and is labelled unverified. Save per-frame detection assignments and unmatched instances so missing real people remain visible. Include the new helper in paid-attempt fingerprints.
Validation: 2,587 Python tests passed; focused matching/accounting tests, Ruff, formatting and diff checks passed. The actual helper preserves all 240 retained elevator and 75 hall-walk associations; a partial city replay retains a continuous 94-sample supported path. These replays omit discarded/raw detections and do not establish all-person recall. Live changed-code trial completed: the main walker now has one continuous ID over all 94 samples, with no duplicate instance assignments; inspected beginning/middle/end overlays no longer track lamps. Background pedestrians still lack poses and 202 unmatched instance observations include real pedestrians and pictured people on signage. Overall all-person acceptance remains failed; no avatar/world was generated. Worker compute estimate $0.0542 excludes startup/builds. Stacked on #131.