Repository navigation
feat(tracking): add opt-in calibrated pose crop recovery - #133
Draft
austinjiann wants to merge 1 commit into
Draft
austinjiann wants to merge 1 commit into
austinjiann wants to merge 1 commit into
Conversation
This was referenced Nov 21, 2026
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Small people can lose detail when a 4K source frame is resized for the pose estimator. This adds an explicit, default-off
--pose-cropsrecovery pass for independent person masks left unmatched by the full-frame pass.Each crop keeps calibrated camera rays, maps predictions back to source pixels, and can claim only its uniquely best-supported target mask. Established full-frame matches remain intact. Competing-person and equal-support cases are refused; retained crop tensors move to CPU before another inference. Audit records preserve crop bounds, origins and counts. The flag and helper code participate in paid-attempt fingerprints; retry limits are unchanged.
Validation: 118 focused geometry/submission/fingerprint tests passed before review corrections; 72 focused tests passed after the competing-instance and GPU-storage corrections. Ruff and
bun run format:checkpass. Full Python suite passed: 2,641 tests (collected before the two additional overlap regressions); the 72-test post-review run includes both new regressions. CI on the committed head is pending. Independent static integration review completed.Draft: the city GPU trial (with reviewed advertisement exclusions from stacked #134) failed all-person acceptance. 186 crop calls returned poses on67calls; five poses met mask support, but all scored0.160–0.220 below the unchanged0.30 track-start threshold. No additional person track survived. Main94/94samples and unique assignments were preserved, but source overlays still miss background pedestrians and show the main body misalignment. Keep off by default; do not repeat this recipe unchanged or present supported-pose counts as tracks.
Worker-reported compute estimate:$0.04697 (112.78seconds), not an invoice. This consumed ordinary tracking attempt3of3 under the existing budget. No avatars, cleaning, world generation or new playable scene followed. All five CI checks passed on this PR head; no merge or quality promotion.