2.5_possible_fire_final

Documentation publique

This update retains all 38,965 original native images and their original Parquet files, image bytes, annotations and splits. It adds 64,374 new unique images, giving 103,339 unique images. The 10 shared images retain their original annotations, including the existing human decisions. Alternate corrected-source records are retained in provenance.

The 24.7 GB original release remains present. This is an additive update; the separate 6.38 GB corrected-input package is not a replacement for that original release. No original image is removed. The separate repository fireviewer/v2.5-box-precise-and-multi-source-suplement remains unchanged at 99f10f68a4550e1ed9be80fc3bb2182db78a6f24.

Load the archival corpus

corpus preserves source objects and task coverage. It is a mixed-task archival view. Use the task-specific COCO files in annotations/merged/ for training rather than treating unannotated tasks as negatives. The original files and their original labels are available in originalv25; historical metadata and the original materializer are in historical/original-v25-1791ffca572e/.

Annotation tasks

firesmoke: category 1 = visible smoke; category 2 = fire, including the original explicit visible-flame labels. The source object label distinguishes flamevisible from the coarser fire category. A generic source fire box is not newly certified as visible flame. visibleflame retains only explicit source flame labels and excludes coarse-fire-only rows. probablefire: category 3; original semantic boxes are mapped individually to proposal boxes, with no global union and no six-box limit.

Original category IDs 1/2 and original source-point provenance are retained. The archival view additionally uses category 3 for probable-fire annotations. An image annotated only for probable fire is excluded from the fire/smoke task. Synthetic flame crops are included only in their eligible task. Full source point annotations remain in the retained source annotation files.

Native pixels, splits and checks

Image bytes are preserved. COCO boxes use raw image pixels; decode without automatic EXIF rotation. Added oriented-coordinate boxes were mapped back to raw coordinates where needed. No resizing or lossy recompression is applied. Original splits and known groups remain fixed. Exact-image hashes are unique and no known group crosses splits. Related-scene and near-duplicate review across the combined sources is not exhaustive.

The 64,384-image correction input was fully decoded and its source COCO annotations reconciled before this merge. Every added image is reread and SHA-256 checked from its staged Parquet. Preservation of the original images is established by leaving all 264 original media shards unchanged and by the complete original hash inventory. No new exhaustive semantic review of the original corpus is claimed.

The 125 synthetic scenes/crops retain their source status pending human acceptance. Source-declared negatives retain their actual authority; absence of a box alone does not create human verification. No inference or training is run during the merge.

Materialize a task

The helper downloads the selected task's exact COCO export and materializes the eligible original and added rows with native image hashes. Use --task visibleflame or --task probablefire for the other tasks. Use validation for model selection and keep test separate.

Provenance and rights

Original source notices, negative-corpus records, historical exposure flags and rights remain in the repository. The corrected source annotations and merge receipts are in provenance/merge-20261024/. Original licences remain source-specific; this update grants no new rights or blanket licence. The previous dataset card is preserved in the historical folder, including its full original-source and association statements.

Merge parent: 1791ffca572e7c2e3dc6b11703621d3b0c2beb64. Complete image preservation and file inventories are recorded in manifest.json and provenance/merge-20261024/image-preservation.json.