Struct EvalGrid 

pub struct EvalGrid {
    pub eval_imgs: Vec<Option<Box<PerImageEval>>>,
    pub eval_imgs_meta: Vec<Option<Box<EvalImageMeta>>>,
    pub n_categories: usize,
    pub n_area_ranges: usize,
    pub n_images: usize,
    pub retained_ious: Option<RetainedIous>,
}

Output of evaluate_bbox / evaluate_segm / evaluate_boundary — the flat (K, A, I) grid of PerImageEval cells the accumulator consumes, plus the dimensions needed to construct crate::accumulate::AccumulateParams.

Fields

eval_imgs: Vec<Option<Box<PerImageEval>>>

Some(cell) per (k, a, i) triple where the cell ran; None where pycocotools would emit None (image absent from detections, no GTs and no DTs in the cell). Layout is K-major, then A, then I — eval_imgs[k * A * I + a * I + i].

Cells are heap-boxed: Option<Box<PerImageEval>> is 8 bytes (Box’s NonNull niche absorbs the discriminant), so the dense n_categories * n_area_ranges * n_images grid only pays for a pointer per slot at zero-init time. On val2017 (1.6M slots, 14k populated) this drops the upfront alloc from 268 MB to 12.8 MB and the zero-init from ~120 ms to ~5 ms — see docs/engineering/benchmarking/2026-05-bbox-cdf.md.

eval_imgs_meta: Vec<Option<Box<EvalImageMeta>>>

Pycocotools-shaped bookkeeping for each populated cell (same [k][a][i] layout as eval_imgs; None wherever eval_imgs is None). Boxed for the same reason as eval_imgs.

n_categories: usize

K axis size: the number of categories used for evaluation, or 1 when use_cats=false.

n_area_ranges: usize

A axis size: equal to params.area_ranges.len().

n_images: usize

I axis size: number of images iterated over (every image in the GT dataset, in deterministic id-ascending order).

retained_ious: Option<RetainedIous>

Per-(category, image) IoU matrices retained when the caller passed EvaluateParams::retain_iou = true. None on the default no-retention path; one discriminant byte wide there.

Implementations

impl EvalGrid

pub fn cell(&self, k: usize, a: usize, i: usize) -> Option<&PerImageEval>

Cell at (category_index, area_index, image_index). Returns None when the indices are in bounds but no cell ran (image absent from detections, or no GTs and no DTs in the cell); returns None for out-of-bounds indices as well.

pub fn cell_meta(&self, k: usize, a: usize, i: usize) -> Option<&EvalImageMeta>

Pycocotools-shaped bookkeeping at (category_index, area_index, image_index). None exactly when EvalGrid::cell is None.

Trait Implementations

impl Clone for EvalGrid

fn clone(&self) -> EvalGrid

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for EvalGrid

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.