auto-safe/ contains approved model proposals and unanimous negative GT. auto-bonzai/ contains Bonsai corrections confirmed with high confidence. manuelle/ contains annotations accepted or corrected by a human. global/ contains distant views with visible fire or smoke outside forest-wildfire scenes. wildfire/ contains distant views with visible forest wildfire. Flame boxes stay small; a distant smoke plume may cover a large area. gold/ is the human calibration holdout; it is excluded from training. Some gold rows provide presence and origin points without boxes; use goldscope. Images and source licences are embedded or recorded per row; there is no blanket image licence. See audit/lots/ for…
Publicly accessible
other
Versioned detector-training corpus. Current V1 repository contents: 87,568 rows, split 53,952 train / 12,787 validation / 20,829 test. The corpus stores source-specific provenance, hashes, grouping/de-duplication information, validation status and annotation metadata. It preserves the V1 training reference and provenance used by earlier strict FireViewer detector releases; each model receipt identifies the snapshot actually used. There is no single licence covering all source images. Keep repository-level license: other and preserve the source licence/attribution recorded for every sample. FireViewer publication must not overwrite source obligations. This corpus is a training/research…
Publicly accessible
other
10K<n<100K
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. 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…
Publicly accessible
other
10K<n<100K
A versioned negative-image resource for fire/smoke detector training and false-positive analysis. The current publication contains 12,653 native source-labelled negatives and 195 retained JPEG candidates, in two separate configurations. The repository name does not establish that every image is a very hard negative. The native subset has no individual visual review or difficulty rating. Existing source classification annotations were the admission basis and were not changed into fresh manual annotations. These identifiers are the source namespaces recorded in the retained provenance, not newly inferred labels or authorizations. See source records and the For the separate JPEG collection…
Publicly accessible
other
10K<n<100K
Versioned geographic map packages and tiled viewer assets for FireViewer. The generic producer is the separate UWD Map Builder; historical FireViewer spatial identifiers remain part of the published compatibility and provenance contracts. These are geographic resources, not a conventional image-training dataset. Map geometry, semantic placements and display assets must remain distinct from dated fire observations, incident perimeters, simulations and model predictions. The current repository contains three runtime/viewer-tiled/catalog.json files: Each catalogue records its bounds, local origin, metre units, asset paths and hashes. Viewer coordinates use a local glTF Y-up frame; preserve the…
Publicly accessible
other
