v2.5-specialisation-visible-fire
Public dataset for object detection, documented on the Hub.
Open the public sourcePublic dataset for object detection, documented on the Hub.
- Downloads
- 0
- Updated
- Oct 05, 26
- Size
- 100K–1M
- Source
- Hugging Face
- Status
- Connected source
Documentation and scope
Structured content from the public source documentation.
Public documentation
RF-DETR visibleflame detection on real probablefire ROIs, with train-only missed-GT jitter and broad context. BBox coordinates are xywh in crop pixels. Empty objects are negative crops; tiny is computed from original-image area.
Source corpus: FireViewer v2.5. Source checkpoint SHA256: 05e431638e3daf9c81c78860ad26b8b65c52765b9d934666d9d4d7da66eb565d. Proposal cache: a5cd7865f2ff51dcb80618369c20cb356bb1fcf642b3ff0794761fc5b8359529.
Train: 137102 ROI rows. Validation: 14822 ROI rows, proposals only. Test remains locked and is not published here during model selection. All crops of one original image remain in its original split. Exact-byte and decoded-image duplicate leakage checks passed in the parent preparation. Event/scene-level independence has not been fully assessed without event metadata; source annotation omissions remain a limitation.
A record traceable to its source
Editorial information, metadata and documentation remain connected to the original public repository.
- Public identifier
- fireviewer/v2.5-specialisation-visible-fire
- Latest activity
- October 5, 2026