v2.5-smoke-from-fire-probability

Quick card

Field	Reference
Use	Study fire-associated smoke compatibility for ROIs without confirmed flame.
Expected input	RGB ROI; binary labels and recorded calibration groups.
Example	A smoke-like ROI → a compatibility label; no point or mask target.
Limits	50,183 train / 12,106 validation public ROIs; no test export. Compatibility is not ignition-origin evidence.
Contract	Schema and calibration groups · Dataset guide

Choose a variant · Resource navigation · Update log

dataset and retained classifier

Binary fire-associated smoke compatibility dataset: 50,183 train / 12,106 validation ROIs, no test export. Linked calibrated classifier: dinov2-small-fire-associated-smoke-v2.5. This is the redistributable subset of the retained run corpus, not its complete calibration/audit data. No smoke points or masks are part of this release.

Positive label=1 (fire_associated_smoke_compatible) means smoke supported as fire-associated by known fire-corpus provenance or visible flame elsewhere in the source image. It does not prove ignition origin, physical danger or aerosol concentration. Positive crops containing any flame fragment and crops with unsupported association are excluded. label=0 (non_fire_or_false_positive) comes from admissible negative proposal regions; uncertainty is held aside, never forced into negatives. The ambiguity interval is an inference decision after calibration, not a training class.

Train combines predominantly real probable-fire predictions with train-only GT fallback; validation is proposals-only. Geometry: 45% context per side and minimum side 160, with exact ROI recorded per row; model input is a bicubic 448×448 resize. Dataset membership is not conditioned on another model predicting no flame. The no-confirmed-flame gate applies only to the operational cascade.

Schema and calibration groups

Embedded image bytes and hash, sample ID, split, binary smoke label, original image hash/name/license, original dimensions, ROI xyxy, proposal origin/score and source checkpoint hash. Preserve all fields; class membership is crop-specific, while grouping is by original image. Model validation uses deterministic original-image-SHA groups (seed 42) with 40% selection / 20% temperature / 20% thresholds / 20% audit. Do not randomly distribute sibling crops among those folds. Refit thresholds on a new domain; the published independent audit concerns the larger retained validation pool, not an automatic score for this public subset.

Frozen Stage-1 checkpoint SHA-256 05e431638e3daf9c81c78860ad26b8b65c52765b9d934666d9d4d7da66eb565d; proposal cache key a5cd7865f2ff51dcb80618369c20cb356bb1fcf642b3ff0794761fc5b8359529. The existing publication-receipt.json and data manifests retain byte hashes and counts; this card update does not reexport or alter those payloads.

Provenance, splits and rights

EN. Derived from v2.5-box-precise-and-multi-source-suplement, original payload revision 5e268f93fb6fad11262123a76bd3579addcf8ea4. Image membership follows the parent's split and original-image SHA-256; every crop of one image stays in its split. Exact-image and decoded-image leakage checks passed upstream. Related frames/events may still occur across splits because complete incident metadata are unavailable. Do not claim event-independent evaluation. The archive fingerprint is 7f0a7d63b9c2b308f786f2b564ef3f1f1c9fc07926d3277d7579c6599a2f3d65.

Only source rows with recorded public redistribution permission are exported. MyData_Fire (Private) and With-Augmentation-Big-1 (undocumented redistribution permission) are excluded. Images retain source-specific licenses: FASDD CC-BY-SA-4.0, Pyro-SDIS Apache-2.0 and recorded licenses for the local/Roboflow additions. license: other is intentional; derivative annotations and packaging do not grant new image rights. Retain source_license, LICENSE, source notices and HF source rights. The notices include source URLs and attribution; they are the authority for reuse, not this summary.

Appropriate use and limitations

Research on visible fire perception, domain-specific detection and localization. No exhaustive human QA, complete incident coverage, demographic/geographic balance or scientific claim about physical fire origin is established. Source boxes can be incomplete or noisy; chromatic heuristics can select luminous non-flame texture. Reevaluate on independently labeled scenes before deployment. TEST was not consulted for point generation, selection or this release.

References

• DINOv2, Oquab et al..
• SAM 2, Ravi et al., official SAM 2 code.
• FASDD, Wang et al., Science Data Bank.
• Pyro-SDIS, Pyronear and partner SDIS.

Likes are always welcome! ❤️ If these resources help you, a little heart makes our day and helps others discover the work. Thank you!