Review Topic Prediction Integrity

Review topic: Prediction, coordinate & pre/post-processing integrity

When this applies

Content trigger (not one directory). Load when a diff does any of:

• Edits coordinate transforms / recovery: sv_detections_to_root_coordinates, scale_sv_detections, attach_parent_coordinates_to_detections, attach_parents_coordinates_to_sv_detections, move_boxes/move_masks, or any xyxy += / * scale / shift arithmetic (inference/core/workflows/core_steps/common/utils.py).
• Touches crop / stitch / slice / perspective blocks: transformations/dynamic_crop, transformations/stitch_images, transformations/stitch_ocr_detections, transformations/perspective_correction, transformations/dynamic_zones, fusion/detections_stitch.
• Changes mask representation: binary mask ↔ polygon ↔ RLE, mask_to_polygon, rle_masks_to_polygons, coco_rle_masks_to_numpy_mask, supported_mask_formats, enforce_dense_masks_in_inference_models, binarization_threshold.
• Touches keypoints: parsing into sv.Detections, keypoint xy/confidence/class arrays, keypoint tensor slice offsets in model post-processing, keypoint visualization/edges.
• Edits model pre/post-processing in inference_models/: normalization, resize/letterbox, BGR/RGB, sigmoid/softmax, NMS params, class/background-index handling, tensor slice layout.
• Changes prediction serialization / response shape: serializers.py, deserializers.py, prediction entity models, block outputs/manifest keys.
• Bumps supervision — its internals (indexing, mask_to_polygons, annotators, edge maps) are load-bearing here.

Review checklist

BLOCK — must be fixed before merge

• All geometry channels move together. A transform that edits xyxy must apply the same shift/scale to mask, data[POLYGON_KEY_IN_SV_DETECTIONS], and keypoints_xy when present. A channel left in a different frame ships wrong results with green happy-path tests (see Standards §1).
• Manifest-declared output keys emitted on every branch. Empty / zero-area / no-detections early returns still emit all declared keys (see Standards §2).
• Backend parity. A post-processing change in one backend is applied to every sibling backend; threshold/sigmoid/slice-offset/index drift changes real users' scores or classes (see Standards §4, §5).
• Serialization contract intact. Field names, types, and rounding downstream expects (predictions, points, keypoints, mask/rle, class, confidence) are preserved; per-keypoint class/class_id/confidence/x/y all present (see Standards §3).
• No class/background-index off-by-one. Class remapping and background index verified against the reference implementation (see Standards §5).

FLAG — reviewer should raise

• Flag parity holds. No numpy-path behavior edit rides a tensor-sibling PR; the same workflow serializes to the same shape in both flag directions unless the divergence is a documented decision; metadata merges preserve per-row class names under colliding class ids (see Standards §8).
• Threshold / activation change is gated. New default threshold, sigmoid-vs-raw-logits, or column-slice change is justified against the reference and made configurable if it shifts scores (see Standards §5).
• No ragged object-dtype arrays. New sv.Detections data arrays are fixed-shape numeric; keypoints padded to uniform length (see Standards §6).
• RLE ↔ dense honored per consumer. Mask-format seam not broken for a downstream consumer (see Standards §3).
• Polygon vertex source. A stored precise polygon is preferred over mask_to_polygon regeneration; int rounding present (see Standards §3).

NIT

• BGR/RGB channel order and PIL-vs-CV2 resize path match the reference decode (see Standards §5).
• supervision bump: annotators, mask_to_polygons, keypoint edge maps, and indexing still behave (see Standards §7).

Not blocking — do NOT demand

• A missing empty/zero-area guard on a branch that is currently unreachable given the block's manifest (note it, don't block).
• Rounding/precision drift that stays within documented numerical tolerance.
• Cosmetic annotator drift from a supervision bump that does not change geometry, scores, or serialized shape.
• Polygon simplification that loses vertices without affecting boxes or class labels.
• Requiring config-gating for a threshold change that provably reproduces the reference output (the change restores parity rather than shifting it — e.g. the YOLO-ultralytics mask default mask_binarization_threshold 0.0→0.5 set via INFERENCE_MODELS_YOLO_ULTRALYTICS_DEFAULT_MASK_BINARIZATION_THRESHOLD, #2212).

Standards

1. Coordinate frames stay consistent. Every geometry channel of a detection — xyxy, mask, data[POLYGON_KEY_IN_SV_DETECTIONS], keypoints_xy — is transformed together by the same shift/scale. sv_detections_to_root_coordinates and scale_sv_detections in common/utils.py are the canonical seam: both apply the shift/scale to POLYGON_KEY_IN_SV_DETECTIONS after xyxy/mask (polygon shift in root recovery #2473; polygon scale #1268). Failure mode: masks/polygons drawn or uploaded at the wrong location, invisible to a test that only checks xyxy.

2. Response shape is a contract. Block outputs match the manifest's declared keys on every branch, including empty/zero-area. Guard len(detections) == 0 before any index-0 ([0]) access into metadata arrays. (dynamic_crop's early return emitted {"crops": None} but omitted "predictions" — #2346.)

3. Serialization round-trips the promised representation. serialise_sv_detections/serialise_rle_sv_detections/mask_to_polygon in common/serializers.py are the response-shape contract: masks emit the polygon/RLE the schema promises with int rounding (.astype(float).round().astype(int).tolist(), #1236) and prefer a stored POLYGON_KEY_IN_SV_DETECTIONS over regenerating via mask_to_polygon; keypoints emit class, class_id, confidence, x, y per point. RLE ↔ dense mask abstraction lives in inference_models/models/base/instance_segmentation.py (supported_mask_formats, coco_rle_masks_to_numpy_mask); the enforce_dense_masks_in_inference_models toggle is a manifest bool field on the instance-segmentation v1/v2 blocks (core_steps/models/roboflow/instance_segmentation/{v1,v2}.py), threaded into the request — not an adapter function (#2384, #2260, #2484).

4. Reference-backend parity. ONNX / TRT / TorchScript paths and legacy-inference vs inference_models paths produce the same predictions within numerical tolerance. When one backend's post-processing changes, every sibling changes too — e.g. the keypoint slice offset fix touched yolov8_key_points_detection_onnx.py, _trt.py, _torch_script.py together (#1626). Failure mode: a model "works" but boxes/scores subtly differ per backend.

5. Threshold / activation / index conventions justified against the reference. New default threshold, sigmoid vs raw logits, background/class index offset, or tensor column slice must match the reference implementation and be config-gated if it shifts scores. The keypoint slice is fixed at image_detections[:, 6:] in run_nms_for_key_points_detection (inference_models/models/common/roboflow/post_processing.py) — not 5 + num_classes (#1626). The YOLO-ultralytics mask default mask_binarization_threshold was set to 0.5 via env const INFERENCE_MODELS_YOLO_ULTRALYTICS_DEFAULT_MASK_BINARIZATION_THRESHOLD (#2212, #2217); note align_instance_segmentation_results's own binarization_threshold parameter default stays 0.0. Class remapping / background-index off-by-one bit RF-DETR seg (#2075, #1619, #1920, #1590). Perspective anchor/extension math (#1234, #1287, #1310, #972). Preprocessing must not silently swap BGR/RGB or introduce PIL-vs-CV2 resize drift.

6. Array shape/dtype discipline in sv.Detections. Ragged object-dtype arrays break supervision's indexing/comparison. add_inference_keypoints_to_sv_detections in common/utils.py pads keypoints to fixed-shape numeric arrays (padded_xy/padded_conf/padded_class_id, uniform max length) rather than dtype=object (#2170).

7. supervision bumps are load-bearing. On any version change, verify annotators, mask_to_polygons, keypoint edge maps, and indexing still behave (#2467, #1725, #1424/#1425 pin history).

8. Flag parity & native-metadata integrity (#2357). Four checks whenever a diff touches tensor-native prediction handling:

   • Flag-off equals pre-tensor main, byte for byte. Any numpy-path behavior edit riding a tensor PR — even a genuine bug fix — is a finding: it must ship as its own change, not as a silent flag-off divergence (semantic-seg confidence-mask relocation lesson: reverted and re-landed separately).
   • Cross-flag payload parity. The same workflow serialized under both flag directions should produce the same shape; kind-order divergence between siblings flips mask encodings (points vs rle_mask) per deployment flag. When a payload difference is a deliberate contract, it must be a documented decision, not a manifest accident.
   • Class-name integrity in native metadata. Native detections carry a per-image class_id→name map in image_metadata plus per-box name overrides in bboxes_metadata; merging/concatenating detections from models with COLLIDING class ids is last-wins on the map — rows silently adopt another model's class name and downstream class-based logic acts on wrong names (consensus/rollup lesson; a raise_on_class_name_conflict knob exists but defaults to override-with-warning). Review any metadata-merge path for this.
   • Golden-response tolerances are GPU-generation-sensitive. Expected-response fixtures with sub-pixel atol (0.1 px) recorded on one GPU generation fail on another with SYSTEMATIC sub-pixel drift across every test — before calling a regression, check detection counts, row ordering, and classes first; matching structure + consistent tiny deltas = numerics, not a product bug (and a runner-hardware upgrade will require golden/tolerance refresh).

Key files & Reference PRs

• inference/core/workflows/core_steps/common/utils.py — sv_detections_to_root_coordinates, scale_sv_detections, attach_parent_coordinates_to_detections, add_inference_keypoints_to_sv_detections. All geometry channels transformed together + keypoint padding (#2473, #1268, #2170).
• inference/core/workflows/core_steps/common/serializers.py — serialise_sv_detections, serialise_rle_sv_detections, mask_to_polygon. Response-shape/serialization contract (#1236).
• inference/core/workflows/core_steps/fusion/detections_stitch/v1.py — SAHI merge via move_boxes/move_masks + OverlapFilter. Pair with transformations/dynamic_crop/v1.py (crop_image, WorkflowImageData.create_crop in execution_engine/entities/base.py, origin-coordinate bookkeeping) (#2346).
• inference_models/inference_models/models/base/instance_segmentation.py — supported_mask_formats, dense/rle abstraction, coco_rle_masks_to_numpy_mask. The enforce_dense_masks_in_inference_models toggle is a manifest bool field on core_steps/models/roboflow/instance_segmentation/{v1,v2}.py selecting dense vs RLE (#2384, #2260).
• inference_models/inference_models/models/common/roboflow/post_processing.py and inference_models/inference_models/models/yolov8/yolov8_instance_segmentation_{onnx,trt,torch_script}.py — backend-parity reference: run_nms_for_key_points_detection keypoint slice [:, 6:] (#1626), align_instance_segmentation_results (binarization_threshold param default 0.0; the 0.5 YOLO default comes from INFERENCE_MODELS_YOLO_ULTRALYTICS_DEFAULT_MASK_BINARIZATION_THRESHOLD, #2212). One backend changes → all siblings change.
• inference/core/utils/rle_to_polygon.py — rle_masks_to_polygons, COCO/uncompressed counts → polygon; compact-mask ↔ polygon reference.

Severity guidance

• Critical (BLOCK) — silent geometry corruption or parity break shipping wrong results with green happy-path tests: a channel (mask/polygon/keypoints) in the wrong coordinate frame; a backend diverging in scores/classes; class/background-index off-by-one; serialization dropping or mislabeling class/confidence/points.
• High (FLAG) — a manifest-declared output key missing on a reachable branch; ragged object-dtype arrays that break supervision indexing; threshold/activation change altering scores without config gate or reference justification; RLE↔dense mismatch that breaks a consumer.
• Medium (NIT / Not blocking) — rounding within tolerance; empty-guard on an unreachable branch; supervision-bump cosmetic annotator drift; polygon simplification not affecting labels/boxes.