flame_smoke_detection_box_not_precise

Public documentation

A prepared two-class detection dataset with 38,965 unique native images and 50,591 bounding boxes. It combines the retained FireViewer detection source snapshot with the published negative corpus, with reproducible grouped splits and a COCO materializer. The legacy repository identifier is retained; this card describes the v2.5 release recorded in manifest.json.

Splits and labels

Train also includes 504 negative images from the detection source; the negative columns above identify only the two imported negative subsets. Native-negative difficulty is unrated; the JPEGs retain their historical candidate designation.

COCO category IDs are 1: smokevisible and 2: flamevisible. Map them explicitly to internal labels 0/1 when required by a model. Boxes are xywh in original raw image pixels. Negatives have empty annotation lists. Source points remain provenance and are not detection targets or geographic coordinates. See categories.json and case-manifest.jsonl.

Pixel and split policy

Original encoded image bytes are preserved. Decode raw pixels without automatic EXIF rotation; no resizing or lossy recompression was used to build this release. Model preprocessing is a separate consumer responsibility.

Seed 42 and an 80/10/10 target were used, with final proportions adjusted to keep groups intact. 10,086 historically model-exposed images and their detected related groups are confined to train. Grouping uses native and decoded RGB hashes, normalized source origins, known capture groups, conservative FLAME sequence grouping and documented perceptual checks. The split receipt reports zero detected cross-split groups or near-duplicate edges, not exhaustive scene-identity proof.

The exclusions audit records 8 conflicting identical-image cases quarantined and 1 decoded-pixel duplicate removed. Source annotation authority is retained; the release does not claim individual human revalidation of every box. Split receipt · Exclusions.

Materialize native COCO folders

In a Python environment with enough storage, retrieve requirements.txt and preparecoco.py from the same pinned revision and install the requirements:

Bind the retained helper to the current canonical repository identifier:

The helper retains a historical repository alias, currently redirected by the Hub; the explicit binding avoids depending on that alias. It uses eight parallel shard downloads, checks native hashes and metadata, and writes train/, valid/ and test/, each with native images and annotations.coco.json. Use validation for selection and keep test closed until final evaluation.

Release verification, provenance and rights

The retained release-verification record reports complete readback of 264 Parquet files, native image hash verification and COCO consistency for the published release. Its ready: true is a packaging and verification state; it does not qualify annotation semantics or a downstream model. The earlier metadata audit has ready: false because full shard verification was recorded separately. Both original records remain.

manifest.json pins the detection and negative source snapshots. The negative source is FireViewer Negative Images. Detailed source declarations remain in provenance. Mixed source-specific licences and attribution apply; publication grants no new image rights. Respect those terms and the original annotation authority.

This is a research/training dataset, not an incident feed, pointing benchmark or emergency decision system.

Documentation and revision

Card updated in English on 5 October 2026, from the public repository at 57a35e4fafbeb0e7c37ae890e97f336de4e79166 and its retained reports. This update checks documentation, repository metadata and small evidence files; it does not rerun training, inference, dataset payload verification or device qualification. Historical receipts keep their original dates, revisions and scope. Earlier README hashes in artifact manifests refer to those earlier releases.

The association FIRE-VIEWER provides administrative and financial stewardship for resources under its control. Technical governance is maintainer-led. Upstream, source-specific and pre-association rights remain separate; repository placement does not transfer ownership or grant additional rights.

Project documentation · Institutional website. Contact: contact@fire-viewer.fr.