ninja990621 commited on
Commit
d97252e
·
verified ·
1 Parent(s): 0bff595

scorevision: push artifact

Browse files
Files changed (1) hide show
  1. README.md +28 -15
README.md CHANGED
@@ -8,7 +8,7 @@ tags:
8
  - scorevision
9
  ---
10
 
11
- # Room-dirtiness detector (Score / SN44)
12
 
13
  Detector for the Score element `manako/Detect-room-dirtiness-detection`.
14
 
@@ -21,19 +21,32 @@ Detector for the Score element `manako/Detect-room-dirtiness-detection`.
21
  | 4 | stain |
22
 
23
  ## How it runs
24
- - `model.onnx`: YOLO11n detector, 640x640 input, box-only head.
25
- - `miner.py`: letterbox resize, onnxruntime on CPU (2 threads), per-class score thresholds,
26
- per-class NMS. Pure numpy/OpenCV, so the chute and Score's CPU compliance run give the same boxes.
27
-
28
- ## How it was trained
29
- - **Starting point**: the public room-dirtiness model at `danieldushku/room`
30
- (Ultralytics YOLO11n-seg). Its backbone was kept frozen; the neck and box head were fine-tuned.
31
- - **Data**: about 2,000 of our own synthetic CCTV-style indoor scenes made with
32
- [Z-Image-Turbo](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo) (Apache-2.0), labelled with
33
- [SAM3](https://huggingface.co/facebook/sam3) using descriptive prompts for the objects each image was
34
- asked to contain. About 15% are clean scenes with none of the five objects.
35
- - **No Score test images were used for training.**
36
- - **Tuning**: per-class thresholds tuned with Score's own per-image scorer on a held-out split of our data.
 
 
 
 
 
 
 
 
 
 
 
 
 
37
 
38
  ## License
39
- AGPL-3.0, following Ultralytics YOLO and the public model this one starts from.
 
8
  - scorevision
9
  ---
10
 
11
+ # Room-dirtiness detector v2 (Score / SN44)
12
 
13
  Detector for the Score element `manako/Detect-room-dirtiness-detection`.
14
 
 
21
  | 4 | stain |
22
 
23
  ## How it runs
24
+ - `model.onnx` is a YOLO11n detector with a 640x640 input and a box-only head.
25
+ - `miner.py` does a letterbox resize and runs onnxruntime on CPU with 2 threads. It applies a 0.25 score threshold and per-class NMS.
26
+ - Everything is plain numpy/OpenCV, so the chute and Score's CPU compliance run give the same boxes.
27
+
28
+ ## How it was made
29
+ This model is a weight-space ensemble (WiSE-FT) of two public models that share the same starting point:
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 the weights.** Both ONNX files were rebuilt as PyTorch checkpoints that reproduce them exactly, to within 0.0004 px.
37
+ - **Blending.** Every Conv+BN pair was blended in its fused (folded) form, as θ = θ_baseline + 0.75 · (θ_finetune − θ_baseline). The result was then exported at 640 px.
38
+ - **Results.** On 20 publicly released test images, scored against an answer key rebuilt from Score's published per-miner verdicts:
39
+
40
+ | model | score |
41
+ |---|---|
42
+ | baseline | 0.397 |
43
+ | the fine-tune, run at 640 | 0.459 |
44
+ | this blend | 0.481 |
45
+
46
+ Any ratio between 0.6 and 0.9 gives 0.477–0.484. Ratios above 1 are worse.
47
+ - **No Score test images were used for training.** The public test images were used only to pick the ratio and to check the result.
48
+
49
+ Thanks to Manako for the baseline and to Targaryen101 for publishing their fine-tune.
50
 
51
  ## License
52
+ AGPL-3.0. Both source ONNX files carry Ultralytics' AGPL-3.0 notice in their metadata, and this derived model keeps it.