AI & ML interests

fire detection, vision, localization, deterministic perimeter, public source analysis, dataset

Recent Activity

Organization Card

FireViewer · Fire, smoke and visual context

FireViewer

From native images to detection boxes and visual context: documented research resources for studying visible fire and smoke.

Current datasets · Selected models · Choose a resource · Full catalogue

Current datasets

Corpus Images Annotations Start here
Probable fire v3 103,244 One common probable_fire proposal class Schema and loading
Fire & Smoke v3 130,444 160,159 boxes · flame / smoke · six image-level contexts Examples and code

The two corpora have different populations and revisions. Images and annotations use native pixels. Fire & Smoke totals include the documented 14 November additions, including the separate legacy-source agreement import; automated corroboration does not establish a new human review.

What do the annotations look like?

Reference annotations for smoke, flame, and flame with smoke

Orange: flame; teal: smoke. The rectangles reproduce annotations from a human-validated Cadryl import, without a model prediction or added confidence score. This illustrative selection is not an accuracy measurement.

Understand XYWH and XYXY · Draw boxes in Python

Selected native detectors

Model Input Classes Recorded validation result Evidence
YOLO26s · L4 928 px flame, smoke Best AP50–95 35.8649%, E20 Snapshot and graph
YOLO11M · A100 928 px flame, smoke Best AP50–95 39.1811%, E46 Snapshot and graph

YOLO11M is the retained improved checkpoint. YOLO26s is a distinct, smaller-checkpoint candidate (approximately 60.2 MB versus 201.3 MB), not a claim of better accuracy or measured on-device performance. Each score retains its training revision and validation protocol; no new shared benchmark was run.

The completed seven-human-lot YOLO11M continuation did not improve the best score and produced byte-identical best weights. Its redundant repository was retired. Result comparison and exact weight identity.

Image-level context classification

Model Input Output Validation macro F1 Validation accuracy
ConvNeXt-Tiny 920 × 920 Six visual-context categories 76.92% 87.72%

The classifier returns a category for the whole image, without boxes. Per-class results and supports describe agreement with the Luna context labels. These categories are not an ordinal danger scale.

Cadryl conversions and historical resources

Resource Role
YOLO11M Cadryl conversion Central package with graph, configuration, manifest and adaptation contract.
ConvNeXt Cadryl conversion Central package with graph, configuration, manifest and adaptation contract.
Historical LiteRT archive Ten distinct conversion packages with their own contracts and qualification scope.
Historical YOLO11M source checkpoint Preserved source of the Cadryl conversion; distinct weights, not the recommended current detector.
Full catalogue Historical models/corpora and geographic scenes.

The two redirect-only Cadryl repositories were retired after checking that their actual packages remain in the central catalogue. Model weights, image rights and software licences remain separate and source-specific. Technical checks do not qualify an autonomous emergency alarm, device integration or real-world danger assessment.

Figure sources · Update log · Association · Project documentation · contact@fire-viewer.fr

English presentation and model selection updated 14 November 2026.