docs/DATASET_V25.md

FireViewer v2.5 · five related datasets

Presentation · Resource index · Variant guide

Public redistributable releases and the retained training corpus are distinct. ROI counts are not independent incidents. Coordinate systems and original human annotations must be preserved.

Dataset	Role	Published population	Coordinates / targets
v2.5-box-precise-and-multi-source-suplement	Train or inspect probable-fire proposals and the separate fire/smoke annotation view.	64,061 images; separate fire_smoke / probable_fire configurations	COCO xywh in original-image pixels
v2.5-specialisation-visible-fire	Inspect or train visible-flame detection on proposed regions.	137,102 train / 14,822 validation ROIs	Boxes in crop pixels; recorded original-image ROI
v2.5-smoke-from-fire-probability	Study fire-associated smoke compatibility for ROIs without confirmed flame.	50,183 train / 12,106 validation ROIs	Binary compatibility labels; no point/mask targets
v2.5-visible-flame-crop-monopoint	Study one visual anchor per accepted visible-flame crop.	30,460 train / 1,812 validation crops	One pseudo-point: local, normalized and original-image coordinates
v2.5-visible-flame-full-image-multipoint	Study detection and multiple per-box flame anchors on original images.	46,576 train / 7,647 validation images	Original boxes plus optional points and point_valid

Preserve the parent relationship

Proposal and crop releases share parent source identities and partitions. Do not randomly repartition crops from one source image across train/validation/test. Keep the recorded ROI offsets when reprojecting model output. Public subsets do not reproduce the retained larger training corpus or balanced training sampler exactly.

Independent annotations

Keep the original detection boxes even where pseudo-point generation abstains. For a full-image object with point_valid=False, ignore only its point loss; retain its detection target. Pseudo-points are not independent human flame-base ground truth. Smoke compatibility has no spatial point or mask target in this release.

Loading, rights and evaluation

Follow each dataset’s pinned loading example and schema rather than inferring meaning from a filename. Retain source-specific redistribution restrictions, attributions, deduplication and exclusion receipts. No raw dataset payload is reverified by this documentation update.

The crop datasets expose train/validation only as documented. Selected validation, pseudo-label agreement and the external 680-image cascade diagnostic have different populations; none replaces the locked final TEST.