fireviewer/dfine-xlarge-fire-smoke-v2

FireViewer D-FINE XLarge Fire/Smoke v2

This repository contains the immutable output of the second FireViewer D-FINE XLarge fire and smoke training campaign. The model detects two visible classes:

• smoke_visible
• flame_visible

Base model

• Model: ustc-community/dfine-xlarge-obj365
• Revision: a99923ec2727499f769a82edba864299fceaca48
• Architecture: DFineForObjectDetection

Training corpus and method

The frozen corpus contains 170,389 unique JPEG images from four adapted COCO manifests. Twenty exact binary JPEG collisions were removed before training. The split contains 114,855 training images, 30,401 validation images, and 25,133 test images. Split leakage and cross-split near-duplicates were checked by the FireViewer preflight.

Training used three epochs, BF16 precision, batch size 5, gradient accumulation 4, an effective batch size of 20, and multi-scale image sizes 640/768/896/960. The run completed 24,048 optimizer steps on an NVIDIA GeForce RTX 5070 Ti.

Evaluation

Split	Loss	mAP	mAP@50	mAP@75	mAR@100
Validation	3.9466	0.1361	0.2263	0.1369	0.6704
Test	3.7421	0.1342	0.2272	0.1344	0.6657

Per-class test mAP is 0.1664 for flame_visible and 0.1020 for smoke_visible. The complete metrics are provided in all_results.json, validation_results.json, and test_results.json.

Integrity

model.safetensors SHA-256:

7efdcd9fc02c7006d06974a7aa13c03171f5ec7414eb80cbb0193eca860c6329

The base revision, manifest digests, hardware profile, split counts, and training parameters are recorded in training-provenance.json and preflight-report.json.

Intended use and limitations

This model is a FireViewer challenger for offline image and retained-keyframe fire/smoke detection. It is not automatically promoted to production. Operational use requires a separate critical detector benchmark, comparison against the currently pinned D-FINE model, calibration, and human review. The preflight explicitly reports the critical-test gate as missing, so this artifact must not be treated as deployment-ready on training metrics alone.

The model can miss small, distant, obscured, or visually ambiguous fire and smoke. Predictions are evidence candidates and must not be interpreted as a confirmed incident or geographic perimeter.