Struct AnchorGrid 

pub struct AnchorGrid {
    pub cx: Vec<f32>,
    pub cy: Vec<f32>,
    pub strides: Vec<f32>,
    pub n_anchors: usize,
}

Flattened anchor descriptors for all FPN levels.

Anchors are ordered level-by-level (stride 8 first, then 16, then 32), and within each level in row-major order (y, x).

Fields

cx: Vec<f32>

Anchor centre x in input-image pixel coordinates. Shape: [A].

cy: Vec<f32>

Anchor centre y in input-image pixel coordinates. Shape: [A].

strides: Vec<f32>

Stride (scale factor from feature to input). Shape: [A].

n_anchors: usize

Total number of anchors A.

Implementations

impl AnchorGrid

pub fn yolo26(img_h: usize, img_w: usize) -> Self

Build the anchor grid for YOLO26 given the input image dimensions.

img_h and img_w must be divisible by 32.

pub fn new(img_h: usize, img_w: usize, strides: &[usize]) -> Self

Build the anchor grid for arbitrary strides.

pub fn decode_ltrb_to_xywh(&self, ltrb: &[f32]) -> Vec<f32>

Decode raw LTRB predictions into absolute XYWH boxes.

Input ltrb: [4, A] channels-first — (l, t, r, b) distances. Output: [4, A] channels-first — (cx, cy, w, h) in pixel coords.

pub fn decode_backward(&self, d_xywh: &[f32]) -> Vec<f32>

Backward through the LTRB → XYWH decode.

d_xywh: [4, A] gradient w.r.t. decoded (cx, cy, w, h). Returns: [4, A] gradient w.r.t. raw (l, t, r, b) predictions.