Struct CocoDetections 

pub struct CocoDetections { /* private fields */ }

COCO detections collection — flat storage plus (image, category)- and per-image indices for the per-cell gather.

Implementations

impl CocoDetections

pub fn from_json_bytes(bytes: &[u8]) -> Result<Self, EvalError>

Loads detections from the JSON array shape pycocotools’ loadRes consumes (a list of objects with image_id, category_id, bbox, score, optional id).

iscrowd and area fields, if present, are silently dropped: quirks E2/J4 force is_crowd=0 and quirk J3 derives area from bbox.

pub fn from_inputs(inputs: Vec<DetectionInput>) -> Result<Self, EvalError>

Builds a CocoDetections from typed inputs. Auto-assigns ids (quirk J1) for inputs that did not supply one, validates finite scores, and derives areas (quirk J3).

pub fn from_records(records: Vec<CocoDetection>) -> Self

Build from already-resolved records, preserving their ids and fields verbatim. Used by the streaming evaluator to assemble a CocoDetections view across batches at finalize/snapshot time without re-running the auto-id and area-derivation logic in Self::from_inputs.

pub fn detections(&self) -> &[CocoDetection]

Flat slice of every detection.

pub fn indices_for(&self, image: ImageId, cat: CategoryId) -> &[usize]

Indices into Self::detections for one (image, category) cell. Empty slice when the cell is empty (no detections of that category on that image).

pub fn indices_for_image(&self, image: ImageId) -> &[usize]

Indices into Self::detections for every detection on an image, regardless of category. Path used when useCats=false (quirk L4).

pub fn lvis_trim(&self, max_dets: i64) -> CocoDetections

LVIS per-image top-max_dets trim (quirk AC2 of ADR-0026).

Mirrors LVISResults.limit_dets_per_image at lvis/results.py:73-84: groups detections by image_id, sorts each group by score descending (stable — quirk AC4), and keeps the top max_dets across all categories combined. The cross-class consequence (quirk AC3): 250 cat-A + 350 cat-B detections on one image trim to 300 total, not 250 + min(350, 300).

max_dets < 0 (or i64::MIN) disables the trim entirely (quirk AC5, mirroring the if max_dets >= 0 guard at results.py:39-40). max_dets == 0 keeps zero detections — edge case the upstream allows but isn’t useful in practice.

The output preserves DT ids and per-detection fields verbatim; only the membership of the flat detections vector and the per-cell index maps change. The original CocoDetections is untouched (the inner Arc<Vec<CocoDetection>> is not shared with the result — the trim copies the surviving entries into a fresh allocation).

Within each image’s group, ties on score resolve in input order: Rust’s slice::sort_by is stable, matching Python’s sorted(_, reverse=True) Timsort behavior. The fact that the matching path’s argsort_score_desc is also stable (np.argsort(-scores, kind="mergesort"), AC4) is a separate invariant — vernier’s parity claim covers both sites.

Trait Implementations

impl Clone for CocoDetections

fn clone(&self) -> CocoDetections

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for CocoDetections

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

Formats the value using the given formatter.