Module float 

Float-domain NMS and box geometry.

Two groups of primitives, both reusable outside the decoder:

• Suppression — nms_float, nms_class_aware_float, and the _extra variants that carry a per-box payload (mask coefficients, say) through suppression so the survivors stay paired with their data.
• Geometry — intersection_area, box_area, iou_value, and ios_value. The tiled-inference merge in crate::tiling shares these, which is what keeps its IoS metric numerically identical to the IoU used by ordinary NMS.

jaccard and jaccard_batch4 are the threshold-test forms used on the suppression hot path; they answer “do these overlap past iou?” without materializing the ratio.

Functions

box_area

Area of a box, clamped to >= 0 per side. Mirrors ModelPack _areas.

intersection_area

Area of the axis-aligned intersection of two boxes, clamped to >= 0 per side. Mirrors the inter term in ModelPack metrics/tiled.py::_match.

ios_value

Intersection-over-Smaller value in [0, 1]. Mirrors ModelPack _match(metric='ios'): inter / max(min(area_a, area_b), 1e-9). Used by the tiled-detection merge: a seam-split object has low IoU but high IoS.

iou_value

Intersection-over-Union value in [0, 1]. Mirrors ModelPack _match(metric='iou'): inter / max(area_a + area_b - inter, 1e-9).

jaccard

Returns true if the IOU of the given bounding boxes is greater than the iou threshold

jaccard_batch4

Batch IoU check: test one reference box a against 4 candidate boxes.

nms_class_aware_float

Class-aware NMS: only suppress boxes with the same label.

nms_extra_class_aware_float

Class-aware NMS with extra data: only suppress boxes with the same label.

nms_extra_float

Uses NMS to filter boxes based on the score and iou. Sorts boxes by score, then greedily selects a subset of boxes in descending order of score.

nms_float

Uses NMS to filter boxes based on the score and iou. Sorts boxes by score, then greedily selects a subset of boxes in descending order of score.

postprocess_boxes_float

Post processes boxes and scores tensors into detection boxes, filtering out any boxes below the score threshold. The boxes tensor is converted to XYXY using the given BBoxTypeTrait. The order of the boxes is preserved.

postprocess_boxes_index_float

Post processes boxes and scores tensors into detection boxes, filtering out any boxes below the score threshold. The boxes tensor is converted to XYXY using the given BBoxTypeTrait. The order of the boxes is preserved.

postprocess_boxes_multilabel_index_float

Multi-label variant of postprocess_boxes_index_float.