v2.5-visible-flame-full-image-multipoint

English — full images with multiple flame points

46,576 train / 7,647 validation images, original archive bytes and dimensions. Flame boxes: 31,908 / 1,950. Valid points: 30,460 / 1,812. Missing point targets: 1,448 / 138. Positive images: 13,199 / 578; remaining 33,377 / 7,069 images are negative for annotated visible flame and may contain smoke. Maximum annotated flame boxes per image: 34 train / 22 validation. These are original-image counts, not ROI counts.

All original flame boxes are retained even when pseudo-point generation abstains. Such objects carry pointvalid=False and pointxy=None; ignore their point loss, retain their detection target. Never substitute (0,0), bbox centre or background. This dataset is an image-level reconstruction of v2.5-visible-flame-crop-monopoint. It supports future multipoint detection research; no satisfactory D-FINE multipoint checkpoint is published. The usable baseline is rfdetr-medium-visible-flame-v2.5 plus dinov2-small-visible-flame-monopoint-v2.5 applied per detection.

Français — images entières multipoints

46 576 images train et 7 647 validation, avec 31 908 / 1 950 boîtes flamme et 30 460 / 1 812 points valides. Les 1 448 / 138 cibles de point manquantes restent nulles : ignorer seulement leur loss de point et conserver la boîte comme cible de détection. Les images sans flamme annotée restent des négatifs utiles, même si elles contiennent de la fumée. Le corpus sert au multipoint futur ; aucun D-FINE multipoint satisfaisant n'est annoncé.

Row schema and independent annotations / Schéma et annotations séparées

image, original-image SHA ID, source and rights metadata, width/height, pseudo-label flags; objects is a dictionary of aligned lists, not a list of dictionaries: bboxxyxy, pointxy, pointvalid, sourceboxindex, pointsampleid. Each valid point has two floats; each unavailable point is null. Empty objects lists mean annotated-negative images. Boxes and points use original-image pixels. COCO detection annotations live in annotations/ .boxes.coco.json with category ID 1=visibleflame and xywh boxes. Point targets are a separate annotations/ .points.json, joined by annotation/image ID; they are not disguised as COCO boxes.

objects contient des listes alignées de boîtes xyxy, points et validité. Les cibles nulles sont explicites. Les fichiers COCO de détection utilisent xywh et catégorie 1 ; les points sont dans un fichier distinct, joint par identifiants. Cette séparation empêche de traiter un point ignoré comme un négatif de détection.

Loading and structural QA / Chargement et QA structurelle

Some clients expose an embedded image as {'bytes': ..., 'path': ...}; cast to datasets.Image() or decode the bytes with PIL explicitly. No custom loading script or automatic remote code is required. DATAFILESMANIFEST.json records payload SHA-256 and size; release.json records parent revision, split counts and generator identity. Every original image hash/dimension, crop bounds, valid point range and reprojection was checked. Parquet shards were reread after writing, and Drive copies were hash-verified. These are structural checks, not a semantic human audit.

Selon le client, l'image embarquée est un objet Image ou un dictionnaire d'octets : décodez-le explicitement. Les manifestes conservent tailles et hashes. Les contrôles de coordonnées, d'octets et de copie Drive ne remplacent pas une revue humaine des points.

Annotation method / Méthode d'annotation

EN. Points are deterministic pseudo-labels called flamebaseanchor. SAM 2.1 Hiera Large is prompted inside each tight original flame-box crop, never on the complete image. Integer crops use floor for the top-left and ceil for the bottom-right, clamped to image dimensions, without context. The highest SAM predicted-IoU candidate mask is used; that score is uncalibrated. The initial target is the mask-bounding-box horizontal centre and y = mask centre + 0.35 × mask height, i.e. 85% from its top. The point is moved to an allowed flame-colour region near that target: HSV red/orange/yellow (hue 0–65 or 350–360, saturation ≥0.35, value ≥0.35), or luminous white (saturation ≤0.22, value ≥0.82). Local golden-yellow pixels (35–65°, saturation ≥0.35, value ≥0.65) are preferred.…

FR. Les points sont des pseudo-annotations déterministes flamebaseanchor. SAM 2.1 fonctionne uniquement dans le crop serré de chaque boîte flamme : aucun contexte ajouté, coordonnées arrondies vers l'extérieur et limitées à l'image. Le masque au meilleur IoU prédit est retenu, sans interpréter ce score comme une confiance calibrée. La cible initiale est au centre horizontal du masque et à 85% de sa hauteur depuis le sommet. Elle est rapprochée des pixels rouge/orange/jaune/blanc lumineux admissibles, avec préférence pour le jaune doré local et les pixels intenses. La composante connexe ayant la masse d'intensité dominante fournit un centroïde, ensuite projeté dans son cœur robuste. Sans masque ou couleur admissible, la génération s'abstient.

This is an appearance-based anchor convention, not an ignition point, geographic fire origin, physical temperature measurement, or one point per incident. The accepted pilot contained 50 boxes from 28 images, with 48 resolved points; the user accepted the green crosses. That acceptance does not validate every corpus annotation. Every public row remains humanreviewed=False and pointlabelsarepseudo=True.

Provenance, splits and rights / Provenance, partitions et droits

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.

FR. Le corpus dérive du dataset principal, révision des données indiquée ci-dessus. Les crops suivent la partition et l'empreinte de leur image d'origine. Les contrôles de doublons ne suffisent pas à garantir l'indépendance des événements ou des séquences ; les métadonnées d'incidents sont incomplètes.

Only source rows with recorded public redistribution permission are exported. MyDataFire (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 sourcelicense, LICENSE, source notices and HF source rights. The notices include source URLs and attribution; they are the authority for reuse, not this summary.