fire1

Fire / Smoke / Fire-extinguisher detector (YOLO26s + ONNX Runtime)

Open-source ScoreVision miner for the manak0/Detect-fire element.

Contents

• weights.onnx — YOLO26s, imgsz 1280, fp16, end2end ([1,300,6] output), ~19 MB (< 30 MB cap).
• miner.py — ONNX Runtime Miner (per-class thresholds, flip-TTA, smoke merge, color-prior FP filters). Reads class order from ONNX names metadata and remaps to ["fire","smoke","fire extinguisher"].
• chute_config.yml — chute image + node selector (1 GPU, 16 GB).
• class_names.txt — fire / smoke / fire extinguisher.

Classes

{0: fire, 1: smoke, 2: fire extinguisher} (matches the element's objects order).

NOTE — re-tune before relying on the score

The per-class confidence thresholds, rescue bonuses, and color-prior gates in miner.py were inherited from the reference top miner and are tuned to that model's score distribution. This is a different YOLO26s (trained on fire_aug), so re-sweep them against this model to maximize the validator's map50 + false_positive pillars.