Crate yscv_detect 

yscv-detect

Object detection pipeline: YOLOv8/v11 decoding, NMS, heatmap detection, ROI pooling, and anchor generation.

use yscv_detect::*;

let detections = detect_people_from_rgb8(width, height, &rgb_data, 0.5, 3, 0.45, 100)?;
for det in &detections {
    println!("{}: {:.0}% at ({}, {}, {}, {})",
        det.label, det.score * 100.0, det.x, det.y, det.w, det.h);
}

Features

• YOLO decoding: YOLOv8 and YOLOv11 output tensor parsing with letterbox preprocessing
• NMS: standard, soft-NMS, batched (class-aware), with early exit optimization
• Heatmap detection: keypoint-based detection with local maxima suppression
• ROI ops: ROI pooling and bilinear ROI align
• Anchors: multi-scale anchor generation for SSD/Faster R-CNN
• Scratch buffers: zero-alloc detection with reusable scratch objects

Optional Features

[features]
onnx = []  # ONNX model inference via yscv-onnx

Tests

60 tests covering decoding, NMS edge cases, ROI alignment.

Structs

BoundingBox

Axis-aligned box in image coordinates.

Detection

One detection with score and class id.

FrameFaceDetectScratch

Reusable scratch storage for frame-based face detection.

FramePeopleDetectScratch

Reusable scratch storage for frame-based people detection.

HeatmapDetectScratch

Reusable scratch storage for connected-component heatmap detection.

ModelDetectorConfig

Configuration for a model-based detector.

Rgb8FaceDetectScratch

Reusable scratch storage for RGB8 face detection.

Rgb8PeopleDetectScratch

Reusable scratch storage for RGB8 people detection.

YoloConfig

YOLO model configuration.

Enums

DetectError

Constants

CLASS_ID_FACE

CLASS_ID_PERSON

CRATE_ID

Traits

ModelDetector

Abstract interface for model-backed object detectors.

Functions

batched_nms

Per-class (batched) NMS.

coco_labels

Returns the 80 COCO class labels.

decode_yolov8_output

Decode YOLOv8 raw output tensor into detections.

decode_yolov11_output

Decode YOLOv11 output tensor into detections.

detect_faces_from_frame

Heuristic face detector over RGB frames using a skin-probability heatmap.

detect_faces_from_frame_with_scratch

Heuristic face detector over RGB frames with reusable scratch storage.

detect_faces_from_rgb8

Heuristic face detector over raw RGB8 bytes.

detect_faces_from_rgb8_with_scratch

Heuristic face detector over raw RGB8 bytes with reusable scratch storage.

detect_from_heatmap

Connected-component detector over heatmaps [H, W, 1].

detect_from_heatmap_with_scratch

Connected-component detector over heatmaps [H, W, 1] with reusable scratch storage.

detect_people_from_frame

Convenience adapter from frame to heatmap-based detection.

detect_people_from_frame_with_scratch

Convenience adapter from frame to heatmap-based detection with reusable scratch storage.

detect_people_from_rgb8

People detector over raw RGB8 bytes.

detect_people_from_rgb8_with_scratch

People detector over raw RGB8 bytes with reusable scratch storage.

detect_yolov8_from_rgb

Run the full YOLOv8 detection pipeline on an HWC image.

detect_yolov8_onnx

Run YOLOv8 inference using an ONNX model.

generate_anchors

Generates anchor boxes for a feature map.

iou

Computes IoU for two axis-aligned boxes.

letterbox_preprocess

Apply letterbox preprocessing: resize an image to a square with padding.

non_max_suppression

Standard score-sorted NMS.

postprocess_detections

Post-processes raw model output into final detections.

preprocess_rgb8_for_model

Preprocess an RGB8 image for model input.

roi_align

RoI Align: bilinear-interpolation version of RoI pooling (no quantisation).

roi_pool

RoI Pooling: for each RoI, divides the region into an output_size grid and max-pools each cell.

soft_nms

Soft-NMS with Gaussian decay.

yolov8_coco_config

Returns the default YOLOv8 config for the COCO 80-class dataset.

yolov11_coco_config

Default COCO config for YOLOv11 models.