Struct EvalImageMeta 

pub struct EvalImageMeta {
    pub image_id: i64,
    pub category_id: i64,
    pub area_rng: [f64; 2],
    pub max_det: usize,
    pub dt_ids: Vec<i64>,
    pub gt_ids: Vec<i64>,
    pub dt_matches: Array2<i64>,
    pub gt_matches: Array2<i64>,
}

Pycocotools-shaped per-cell bookkeeping that the matching engine strips out when packing PerImageEval. Surfaced separately so the accumulator stays narrow per ADR-0005, and FFI / COCOeval drop-in consumers can reconstruct evalImgs dicts without re-running eval.

All dt_* axes are in score-descending sorted order (stable mergesort, quirk A1); all gt_* axes are in ignore-ascending sorted order (quirk A4). dt_matches and gt_matches carry pycocotools’ value semantics: i64 annotation ids on a hit, 0 on a miss (matching dtm/gtm initialization in cocoeval.py).

Fields

image_id: i64

COCO image id for this cell.

category_id: i64

COCO category id, or COLLAPSED_CATEGORY_SENTINEL when use_cats=false.

area_rng: [f64; 2]

Active area range as [lo, hi], mirroring pycocotools’ aRng.

max_det: usize

max_dets_per_image cap that produced this cell’s DT slice.

dt_ids: Vec<i64>

DT annotation ids in sorted-DT order, length D.

gt_ids: Vec<i64>

GT annotation ids in sorted-GT order, length G.

dt_matches: Array2<i64>

Shape (T, D). GT id matched at (threshold, sorted-DT k), or 0 if unmatched (pycocotools sentinel; safe because COCO ids are >= 1 per spec, and vernier’s auto-id assignment also starts at 1).

gt_matches: Array2<i64>

Shape (T, G). DT id matched at (threshold, sorted-GT k), or 0 if unmatched (same >= 1 invariant as dt_matches).

Trait Implementations

impl Clone for EvalImageMeta

fn clone(&self) -> EvalImageMeta

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for EvalImageMeta

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

Formats the value using the given formatter.