LibreYOLO/coco-val2017-mini500

COCO val2017 mini-500

A frozen, reproducible 500-image subset of COCO val2017 for fast object-detection latency/throughput benchmarking on edge devices (e.g. LibreYOLO on NVIDIA Jetson), where the full 5000-image val set is impractical.

Selection (deterministic)

The 500 images are sorted(COCO.getImgIds())[:500] of the official instances_val2017.json (the 500 lowest image IDs) — identical to running the vision-analysis-benchmark harness with --limit 500.

field	value
images	500
annotations	3541
image_ids SHA-256	74bea76eee276928bb1fb2f3b30e6a6e2760124a1de4931727b50080ccf81951
subset annotations SHA-256	31c8e48b3e0bc957da945ad43697f8d519d7aaeb5971cef19e4cfaa4a8423a56
source instances_val2017.json SHA-256	e8c7f7908f1d7278341fae127d0da654f102f11bd7b21d8aeefa635b8c810b6f

Structure

annotations/instances_val2017_mini500.json   # COCO format, filtered to the 500 images
images/val2017/*.jpg                          # the 500 images
manifest.json                                 # image_ids + file_names + SHA-256 fingerprints

Usage

from pycocotools.coco import COCO
coco = COCO("annotations/instances_val2017_mini500.json")
print(len(coco.getImgIds()))  # 500

Provenance, license & attribution

Derived from COCO 2017 (Lin et al., Microsoft COCO: Common Objects in Context, ECCV 2014) — https://cocodataset.org. Annotations are CC BY 4.0 (COCO Consortium). Images are from Flickr and retain their original licenses as listed in the COCO licenses field (preserved in the annotations file); re-hosted unmodified only to make this benchmark subset turnkey. No new annotations are added — this is a deterministic repackaged slice of the official release.