v2.5-box-precise-and-multi-source-suplement

Public documentation

64,061 public images, from the final deduplicated trainv2.5final1 corpus. 7,574 retained MyDataFire images and all their annotations are excluded because their source metadata says Private. The ongoing training corpus is preserved separately.

A further 12,025 images from sources with undocumented redistribution rights are held outside this public release.

Separate annotation variants

Both variants contain exactly the same images, bytes and train/validation/test membership. Boxes use COCO xywh pixels at original archive dimensions. Fire includes the original flame/fire categories; smoke retains its source category. The probable-fire annotations match the current training annotations for every admitted public image. Identical fire and smoke boxes may be represented twice in the separate-class variant.

Loading

Use preparecoco.py to materialize native images and the selected COCO annotations in a remote CPU runtime. No image resizing or recompression is performed during this publication. Keep test for final evaluation.

Source mapping and checks

The legacy HF source maps 0 to smoke and 1 to fire; local exports have their own namespaces. Publication never uses a universal interpretation of input YOLO IDs. Original Roboflow COCO category names are recovered through the retained image-pairing records. Every published image hash, dimension, final box, split count and Parquet row is checked before upload. See DATAFILESMANIFEST.json, provenance/release-input.json and provenance/source-notices.json.

The title is a version name, not a claim of exhaustive human validation or mathematically exact boundaries. Some source boxes were automatically tightened in the upstream archive. The original authority is retained; source pairing and structural checks do not establish the semantic correctness of every box. Known duplicate groups have been removed; unseen related scenes remain possible.

Sources and rights

There is no blanket licence for all images. Repository metadata uses license: other. Retain the per-row source licence and the source notices. This release grants no new image rights. Private MyDataFire images, labels and source records are not published.

Source attribution and redistribution conditions are retained in provenance/hf-source-rights.json. FASDD images and their adaptations retain CC-BY-SA-4.0; Pyro-SDIS images retain Apache-2.0. Roboflow additions retain their individual recorded licences. Changes in this release: public-source selection, upstream box refinements, strict duplicate removal, source-specific category recovery and packaging of two annotation variants.

Legacy HF annotations are traced to fireviewer/fire-smoke-detection-corpus-v1. Archive SHA-256: 7f0a7d63b9c2b308f786f2b564ef3f1f1c9fc07926d3277d7579c6599a2f3d65. Training-corpus fingerprint: ab4360d4ef3f6e1f419a317b9d8ddff5602c8e3e3394babd1bd0f3faa8916efe.

• hf-v1:fasdd: 29005 images.
• hf-v1:pyro-sdis: 9414 images.
• local-add:Small Fire Smoke OD Dataset.v6i.coco.zip: 1186 images.
• local-add:fire.v1i.coco.zip: 7877 images.
• roboflow:détection-d'incendie-2: 10321 images.
• roboflow:mixfirSmok-12: 2411 images.
• roboflow:smoke-and-fire-detection-22: 3847 images.
• FASDD: Ming Wang, Peng Yue, Liangcun Jiang, Dayu Yu, Tianyu Tuo and Jian Li; FASDD / Science Data Bank. CC-BY-SA-4.0.