scorevision: push artifact
Browse files
README.md
CHANGED
|
@@ -8,7 +8,7 @@ tags:
|
|
| 8 |
- scorevision
|
| 9 |
---
|
| 10 |
|
| 11 |
-
# Room-dirtiness detector
|
| 12 |
|
| 13 |
Detector for the Score element `manako/Detect-room-dirtiness-detection`.
|
| 14 |
|
|
@@ -22,29 +22,32 @@ Detector for the Score element `manako/Detect-room-dirtiness-detection`.
|
|
| 22 |
|
| 23 |
## How it runs
|
| 24 |
- `model.onnx` is a YOLO11n detector with a 640x640 input and a box-only head.
|
| 25 |
-
- `miner.py`
|
| 26 |
-
-
|
|
|
|
| 27 |
|
| 28 |
## How it was made
|
| 29 |
-
|
| 30 |
|
| 31 |
| weight | model | license |
|
| 32 |
|---|---|---|
|
| 33 |
| 0.25 | [manako/Detect-room-dirtiness-detection](https://huggingface.co/manako/Detect-room-dirtiness-detection), the Manako baseline (YOLO11n-seg) | AGPL-3.0 |
|
| 34 |
| 0.75 | [Targaryen101/dirtiness_01](https://huggingface.co/Targaryen101/dirtiness_01), a full fine-tune of that baseline | AGPL-3.0 |
|
| 35 |
|
| 36 |
-
- **Rebuilding
|
| 37 |
-
- **
|
| 38 |
-
- **Results.** On 20 publicly released test images, scored against an answer key rebuilt from Score's published per-miner verdicts:
|
| 39 |
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
| baseline | 0.397 |
|
| 43 |
-
| the fine-tune, run at 640 | 0.459 |
|
| 44 |
-
| this blend | 0.481 |
|
| 45 |
|
| 46 |
-
|
| 47 |
-
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
| 49 |
Thanks to Manako for the baseline and to Targaryen101 for publishing their fine-tune.
|
| 50 |
|
|
|
|
| 8 |
- scorevision
|
| 9 |
---
|
| 10 |
|
| 11 |
+
# Room-dirtiness detector v3 (Score / SN44)
|
| 12 |
|
| 13 |
Detector for the Score element `manako/Detect-room-dirtiness-detection`.
|
| 14 |
|
|
|
|
| 22 |
|
| 23 |
## How it runs
|
| 24 |
- `model.onnx` is a YOLO11n detector with a 640x640 input and a box-only head.
|
| 25 |
+
- `miner.py` letterbox-resizes the frame and runs onnxruntime on CPU with 2 threads. Everything is plain numpy/OpenCV, so the chute and Score's CPU compliance run give the same boxes.
|
| 26 |
+
- Score thresholds are 0.25 for every class except stain, which uses 0.35. Boxes then go through per-class NMS.
|
| 27 |
+
- **One box per object.** A box is dropped when it lies at least 70% inside a bigger box of the same class whose score is at least 0.9 times its own. Score's labeller gives each object a single box, so such a box is usually one piece of a larger object, for example one item of a litter pile.
|
| 28 |
|
| 29 |
## How it was made
|
| 30 |
+
The weights are a weight-space ensemble (WiSE-FT) of two public models that share the same starting point:
|
| 31 |
|
| 32 |
| weight | model | license |
|
| 33 |
|---|---|---|
|
| 34 |
| 0.25 | [manako/Detect-room-dirtiness-detection](https://huggingface.co/manako/Detect-room-dirtiness-detection), the Manako baseline (YOLO11n-seg) | AGPL-3.0 |
|
| 35 |
| 0.75 | [Targaryen101/dirtiness_01](https://huggingface.co/Targaryen101/dirtiness_01), a full fine-tune of that baseline | AGPL-3.0 |
|
| 36 |
|
| 37 |
+
- **Rebuilding and blending.** Both ONNX files were rebuilt as PyTorch checkpoints that reproduce them exactly. Every Conv+BN pair was then blended in its fused (folded) form, as θ = θ_baseline + 0.75 · (θ_finetune − θ_baseline).
|
| 38 |
+
- **Spill row.** In the final classification conv, the row for liquid spillage is taken 50/50 from the baseline and the blend.
|
|
|
|
| 39 |
|
| 40 |
+
## Results
|
| 41 |
+
Scores below use Score's own per-image scorer. The public test images are scored against answer keys rebuilt from Score's published per-miner verdicts.
|
|
|
|
|
|
|
|
|
|
| 42 |
|
| 43 |
+
| | 14 well-pinned public tests | all 31 public tests | 962 of our own generated images (SAM3 labels) |
|
| 44 |
+
|---|---|---|---|
|
| 45 |
+
| baseline, run the way public copies run it | 0.533 | 0.460 | 0.324 |
|
| 46 |
+
| v2 (blend, plain 0.25 threshold) | 0.551 | 0.502 | 0.368 |
|
| 47 |
+
| **v3 (this release)** | **0.567** | **0.515** | **0.373** |
|
| 48 |
+
|
| 49 |
+
- **The one-box-per-object rule** was also checked on the 962 generated images, which had no part in tuning it. It improved 103 images and worsened 14.
|
| 50 |
+
- **No Score test images were used for training.** The public test images were used only to choose the blend ratio and the two thresholds, and to check the result.
|
| 51 |
|
| 52 |
Thanks to Manako for the baseline and to Targaryen101 for publishing their fine-tune.
|
| 53 |
|