manak0/Detect-fire-winner

ScoreVision-Fire — meaculpitt v2.1

SN44 fire-detection miner for the manak0/Detect-fire element.

Pipeline

• Architecture: yolo26n
• Resolution: 1408×768 input → letterbox → 960×960
• Preprocessing: cv2.dnn.blobFromImage (fused C++ resize+normalize+transpose)
• Inference: single-pass FP16 ONNX, NMS baked in
• Output shape: [1, 300, 6] (xyxy, conf, cls)
• Latency: ~35 ms p95 on RTX 4090 (fits the 50 ms gate)

Classes (validator GT order, NOT the published class_names.txt order)

• 0: fire
• 1: smoke
• 2: fire extinguisher

Verified by audit of alfred8995/fire001 (scores 1.00) and navierstocks/fire (scores 0.96): both use [fire, smoke, fire_extinguisher] and the validator's GT order matches. Our model was trained with [fire, fire_ext, smoke]; miner.py applies cls_remap=[0,2,1] to translate model output to validator index.

Training

• 22,796 training images (validator-synth + Simuletic + D-Fire + z5atr, SHA1 deduped)
• 2,532 validation images (random 90/10 split, seed=42)
• 100 epochs, yolo26n, imgsz=960, batch=8, AdamW lr0=0.001 cos_lr
• CCTV augmentation chain (cctv_aug_patch)

Benchmarks

• Broader merged val mAP50: 0.785
• Validator-distribution synth val mAP50: 0.640 (+24.7 pts above 0.393 baseline)
• Per-class on synth val: fire=0.523, fire_extinguisher=0.647, smoke=0.749