run string | input dict | params dict | total_raw_detections int64 | by_class dict | detections_per_frame dict | confidence_histogram dict | kept_at_threshold dict | delta_vs_run2 dict |
|---|---|---|---|---|---|---|---|---|
run3_recall | {
"frames": 5323,
"blocks": 12,
"preprocess": "per-frame CLAHE (clipLimit=2.0, tile=8x8) on dimmed thermal",
"peak_temp_C": "unavailable (8-bit masked JPG, no radiometric scale)"
} | {
"conf": 0.1,
"iou_nms": 0.7,
"agnostic_nms": false,
"imgsz": 1280,
"augment_TTA": "requested=True, EFFECTIVE=False (YOLO26 end-to-end model does not support TTA; Ultralytics reverted to single-scale @ imgsz=1280)",
"crop": "NONE (full frame)",
"post_filters": "NONE",
"nms_note": "YOLO26 is NMS-free en... | 1,463 | {
"open_circuit": 712,
"single_bypassed": 374,
"multi_hotspot": 377
} | {
"frames_total": 5323,
"frames_with_detection": 972,
"frames_zero_detection": 4351,
"frames_single_detection": 658,
"frames_multi_detection": 314,
"pct_single_of_detection_frames": 67.7,
"pct_multi_of_detection_frames": 32.3,
"distribution_ndets_to_nframes": {
"0": 4351,
"1": 658,
"2": 196,... | {
"0.10-0.20": 536,
"0.20-0.30": 247,
"0.30-0.40": 154,
"0.40-0.50": 123,
"0.50-0.60": 91,
"0.60-0.70": 83,
"0.70-0.80": 102,
"0.80-0.90": 117,
"0.90-1.01": 10
} | {
">=0.10": 1463,
">=0.20": 927,
">=0.30": 680
} | {
"raw_dets": "428 -> 1463",
"pct_single_detection_frames": "80.0 -> 67.7",
"note": "run-2 was a ~339-frame set; run-3 is the full 5,323-frame flight-2 set, so absolute raw-det counts are not directly comparable; the single-detection % is the crop-removal signal to watch."
} |
No-Filter Results — Run 3 (recall pass)
Raw, unfiltered detection output from running the
merdiofrivia/YOLO26 thermal solar-PV fault
detector over the full
merdiofrivia/Szaszberek-flight2-masked-thermal
set (5,323 frames / 12 blocks).
All gates OFF — full frame (no crop), conf 0.10, no temperature ceiling, no mask post-filter, no corroboration/≥2-frame gate, no unsnapped-drop. Every raw detection kept; no projection / snapping / dedup.
Inference settings
- Input: per-frame CLAHE (clipLimit 2.0, tile 8×8) on the dimmed masked thermal frames.
- conf=0.10, imgsz=1280, full frame.
augment=Truewas requested but YOLO26 is an end-to-end model → reverted to single-scale.- YOLO26 is NMS-free end-to-end, so iou/agnostic_nms apply no suppression.
Results
- 1,463 raw detections across 972 frames (4,351 frames had zero).
- By class: open_circuit 712 · multi_hotspot 377 · single_bypassed 374 · hotspot 0.
- 67.7% of detection-bearing frames are single-detection (run-2: 80%).
- Kept at threshold: ≥0.10 → 1,463 · ≥0.20 → 927 · ≥0.30 → 680.
Files
detections.csv— one row per raw detection (frame, block, class, conf, x1, y1, x2, y2, peak_temp_C).peak_temp_Cis blank: 8-bit masked JPGs carry no radiometric scale.summary.json— counts, per-frame distribution, confidence histogram, run-2 delta.run.log— full inference log.
- Downloads last month
- 30