model card: v6 edge release results and thresholds
Browse files
README.md
CHANGED
|
@@ -23,7 +23,7 @@ cameras and drones, tuned for FAINT early-stage plumes.
|
|
| 23 |
|---|---|---|---|---|
|
| 24 |
| `expB_y11s_deployed.pt` | YOLO11s @1280 | 9.4 M | pyro-sdis train + D-Fire train | production default of the project service |
|
| 25 |
| `v5bA_y11m_best.pt` | YOLO11m @1280 | 20.1 M | same recipe, retrained (v5b arm A, seed 3) | best fixed-camera model on the frozen test |
|
| 26 |
-
| `
|
| 27 |
|
| 28 |
## Headline numbers (frozen pyro-sdis TEST, matched negative-image FP rate, IoU≥0.4)
|
| 29 |
|
|
@@ -62,3 +62,26 @@ aerial) — see the repo's eval tooling.
|
|
| 62 |
|
| 63 |
AGPL-3.0 (derivatives of Ultralytics YOLO11). Dataset licences remain with
|
| 64 |
their owners.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 23 |
|---|---|---|---|---|
|
| 24 |
| `expB_y11s_deployed.pt` | YOLO11s @1280 | 9.4 M | pyro-sdis train + D-Fire train | production default of the project service |
|
| 25 |
| `v5bA_y11m_best.pt` | YOLO11m @1280 | 20.1 M | same recipe, retrained (v5b arm A, seed 3) | best fixed-camera model on the frozen test |
|
| 26 |
+
| `v6A_y11n/v6A_y11n_best.pt` | YOLO11n @1280 | 2.59 M | pyro-sdis train + D-Fire train (real only) | v6 edge model (+ ONNX / TFLite fp16 / TFLite INT8 3.6 MB) |
|
| 27 |
|
| 28 |
## Headline numbers (frozen pyro-sdis TEST, matched negative-image FP rate, IoU≥0.4)
|
| 29 |
|
|
|
|
| 62 |
|
| 63 |
AGPL-3.0 (derivatives of Ultralytics YOLO11). Dataset licences remain with
|
| 64 |
their owners.
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
## v6 edge release (YOLO11n, real-data-only)
|
| 68 |
+
|
| 69 |
+
`v6A_y11n/` contains the frozen v6 deliverable: PyTorch checkpoint, ONNX,
|
| 70 |
+
TFLite fp16 and TFLite INT8 (3.6 MB), plus `infer.py` — a single-command
|
| 71 |
+
offline image/video inference script (PyTorch / ONNX / TFLite backends) —
|
| 72 |
+
and the export report.
|
| 73 |
+
|
| 74 |
+
Frozen pyro-sdis TEST (matched negative-image FP, IoU≥0.4): R@5% 0.243,
|
| 75 |
+
R@10% 0.344 — the best nano-scale model of the v6 matrix. Deployment
|
| 76 |
+
validation of the INT8 file itself on the project val set: no loss at the
|
| 77 |
+
2%/5% FP operating points, −2.1 pt at 10% (traced to graph conversion, not
|
| 78 |
+
quantization; fp16 TFLite is bit-identical in accuracy). Suggested
|
| 79 |
+
confidence thresholds from the INT8 file's own val curve: 0.73 / 0.68 /
|
| 80 |
+
0.62 for ~2 / 5 / 10 % negative-image FP. CPU x86 median forward latency
|
| 81 |
+
45.8 ms (proxy; on-device latency unmeasured).
|
| 82 |
+
|
| 83 |
+
v6 synthetic-data result (pre-registered, reported as-is): neither
|
| 84 |
+
procedural compositing nor repaired generative image-to-image synthesis
|
| 85 |
+
improved real-data evaluation at matched optimizer steps; seed spread
|
| 86 |
+
(±3 pt) exceeded all between-arm differences. Full protocol and reports in
|
| 87 |
+
the GitHub repo (`reports/v6/`, `reports/v6/repair/`, `reports/v6/sprint/`).
|