YOLOV3Config

Configuration for YOLOv3.

YOLOv3 (Redmon & Farhadi, 2018) uses a Darknet-53 backbone with three detection scales at strides 8, 16, and 32 (P3/P4/P5). Each scale has 3 hand-picked COCO anchors for small, medium, and large objects respectively. YOLOv3-Tiny shares this config with its own table — six anchors, strides 16 and 32 — which yolo_v3_tiny supplies.

Args: num_classes: Number of foreground classes (COCO default = 80). in_channels: Input image channels. anchors: (width, height) anchor priors in pixels — every prior a ground-truth box is matched against. strides: Feature-map stride of each detection scale, fine to coarse (P3, P4, P5 for the full model). score_thresh: Minimum objectness × class score to keep a detection. nms_thresh: IoU threshold for per-class NMS. lambda_coord: Box regression loss weight. lambda_noobj: Objectness loss weight for negative anchors. anchor_masks: Indices into anchors predicted at each scale, in strides order — the cfg's mask = per [yolo] layer. Must hold one mask per stride.