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

Quick card

Field	Reference
Use	Study detection and multiple per-box flame anchors on original images.
Expected input	Original RGB image; all flame boxes retained, with point_valid and optional point_xy.
Example	An object with point_valid=False keeps its detection target and skips only point loss.
Limits	46,576 train / 7,647 validation images. Missing point targets and parent partitions must be preserved.
Contract	Row schema and independent annotations · Dataset guide

Choose a variant · Resource navigation · Update log

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 point_valid=False and point_xy=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.

Row schema and independent annotations

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: bbox_xyxy, point_xy, point_valid, source_box_index, point_sample_id. 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/<split>.boxes.coco.json with category ID 1=visible_flame and xywh boxes. Point targets are a separate annotations/<split>.points.json, joined by annotation/image ID; they are not disguised as COCO boxes.

Loading and structural QA

from datasets import load_dataset
ds = load_dataset('fireviewer/v2.5-visible-flame-full-image-multipoint', split='train', streaming=True)
row = next(iter(ds))

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. DATA_FILES_MANIFEST.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.

Annotation method

EN. Points are deterministic pseudo-labels called flame_base_anchor. 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. Intensity combines 0.7 HSV value and 0.3 linear-RGB luminance; the upper local quartile is retained. An eight-connected component is chosen by weighted intensity mass; its weighted centroid is projected inside a robust component core using the distance transform. Pixel-centre coordinates use +0.5. Empty masks or colour support cause abstention, not fabricated coordinates.

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 human_reviewed=False and point_labels_are_pseudo=True.

SAM source: facebook/sam2.1-hiera-large, revision 665f8e2ad61cf5f53d65644ff27c8ee525124610; weights SHA-256 dc407dce21301fd94abb395c5099b4f2c455fdc8a8f261ac3d0ea6d4cd197230. Corpus key: 06001121391d953d12379a97315089bf273f1ac5e9d21067b061638010d17abd. Point dataset key: 4c1b4ad38ca3c6b8b8120094c6714a81383c57ad8f57690adaff1d38644c2aa7.

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!