Filter Pose Bounding Box

1 VersionUtility

Estimates yaw from a Landmarks2d.Human22 pose using shoulder spread and spine height and emits Bool true when the subject is frontal enough. A BoundingBox anchors the subject. Use to gate face or identity work on frontal poses.

How it fits

Typical backends

Detect human landmarks then gate face or identity work on frontal subjects.

Gate frontal subjects before feeding a multi-object tracker so back-facing detections are dropped.

Caveats

Hard constraintThe yaw formula is wired to the Landmarks2d.Human22 schema using shoulder, spine, and thorax keypoints; a different landmark schema type-checks but produces a meaningless yaw value.
AccuracyYaw assumes an upright subject. Seated, lying, or strongly tilted poses are misclassified because the spine-vertical projection no longer tracks body height.
AccuracyYaw is computed from raw coordinates with no per-keypoint confidence weighting, so occluded shoulders yield an unreliable result; pre-filter on landmark confidence upstream when reliability matters. Degenerate geometry falls back to a neutral mid-value.
Parameter interactionyaw_threshold is direction-agnostic; equal left and right rotations are treated identically and a subject above the threshold is rejected as facing away.
I/O contractThe BoundingBox input is consumed as the subject anchor but does not affect the predicate, which is derived only from the pose; one subject is handled per call, so split multi-person detections upstream.

Versions

c8cbd7dclatestdefault
Automated release
6/27/2026