sp500_yolo_params¶
Model shape and thresholds.
Fill this with sp500_yolo_params_init(). The defaults fit a 640x640, 80-class detection model, so a plain detector needs no field set at all. img_w, img_h, num_classes and max_detections must be in 1..65535.
Attributes¶
| Member | Description |
|---|---|
img_w |
|
img_h |
|
num_classes |
|
max_detections |
|
conf_threshold |
|
iou_threshold |
|
anchors_per_level |
|
anchors |
|
anchors_count |
|
reserved0 |
|
proto |
|
mask_coeffs |
|
proto_h |
|
proto_w |
|
mask_threshold |
|
reserved |
img_w¶
network input width in pixels; default 640
img_h¶
network input height in pixels; default 640
num_classes¶
class count; default 80. Pose heads use 1
max_detections¶
upper bound on returned records; default 100. Also sizes output.records
conf_threshold¶
scores below this are discarded; default 0.25
iou_threshold¶
NMS overlap threshold; default 0.7. Ignored by the top-k families (NMSFREE, DFLFREE)
anchors_per_level¶
anchor boxes per feature level, 1..3; default 3
anchors¶
anchors_per_level * 3 (w,h) pixel pairs, level-major for strides 8, 16 and 32. NULL selects the COCO v5 defaults, which requires anchors_per_level == 3. Read only during the call. Default NULL
anchors_count¶
floats in anchors; must be anchors_per_level * 6 when set, else 0
reserved0¶
must be zero
proto¶
prototype masks, [proto_h][proto_w][mask_coeffs] float
mask_coeffs¶
mask coefficients per detection, typically 32
proto_h¶
prototype height, typically img_h / 4
proto_w¶
prototype width, typically img_w / 4
mask_threshold¶
a mask pixel is set when its logit exceeds this; default 0
reserved¶
must be zero