galthran-wq / face-recognition-service

89%

push

github

Pad square fallback (#5)

* Expose 5-point landmarks in detect/embed/analyze responses

Populates the `landmarks` field on DetectedFace from the kpss returned by
the SCRFD detector, and serializes them as LandmarkPoint objects in the
API response. Enables clients to perform landmark-based checks (e.g.
face orientation sanity) without running their own face detector.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Pad-to-square fallback when detector returns zero faces

RetinaFace anchors miss faces that fill most of the frame. On a miss,
pad the image to a square with a 100px gray border and re-run detection
once. Translate bbox/landmark coordinates back to the original frame
(clipping bboxes to the original dimensions) so callers never see
padded coordinates.

Applied uniformly to detect/embed/analyze and the embed_batch /
analyze_batch paths.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Make pad-fallback border/fill configurable; add review follow-up tests

Review follow-ups on pad-square-fallback:
- Move border_px / fill from module constants to config + provider ctor
  (face_pad_fallback_border_px, face_pad_fallback_fill).
- Add test that alignment receives the padded canvas, not the original
  (regression guard: silently using `img` instead of `working` would
  pass earlier tests).
- Mirror embed_batch fallback test for analyze_batch.
- Add test asserting non-default border/fill flow through to the canvas.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

81 of 85 new or added lines in 5 files covered. (95.29%)

571 of 643 relevant lines covered (88.8%)

0.89 hits per line