GemScan inclusion detectors (ONNX)
YOLO11s and RT-DETR-L fine-tuned to detect inclusions in diamond photos, exported to ONNX
(opset 17, static 1×3×640×640 input, FP32). manifest.json lists each file with its
architecture, input size and class names.
Test split (16 images, 117 inclusions):
| Model | Precision | Recall | mAP50 | mAP50-95 |
|---|---|---|---|---|
| YOLO11s | 0.390 | 0.213 | 0.187 | 0.059 |
| RT-DETR-L | 0.357 | 0.282 | 0.195 | 0.067 |
Trained on Diamond Inclusion (CC BY 4.0) by Diamond Classification Data.
Pre/post-processing that reproduces Ultralytics exactly is in app/inference.py of the source repository.