rfdetr_nano_onnx
ONNX export (FP32) of Roboflow RF-DETR nano, used by the Ipsilon apps for on-device vehicle detection (car, motorcycle, bus, truck) before licence-plate localisation.
- Upstream: Roboflow/rf-detr-nano, code at https://github.com/roboflow/rf-detr
- Licence: Apache-2.0 (same as upstream)
- Provenance: exported from the upstream checkpoint with
export_rfdetr_onnx.py(opset 17, FP32) and checked against the PyTorch model withverify_onnx_accuracy.py(logits and boxes correlation above 0.99).
Files
| File | Content |
|---|---|
rfdetr_nano.onnx |
model, 107,995,247 bytes, SHA-256 ba72a8258f7dc383d0d18246acf26204c65f51cec3e64cd5e07cb57eaff603e1 |
id2label.json |
class index to label (COCO 91-class ids, index 0 is N/A) |
preprocessor_config.json |
normalisation constants |
Input and output
- Input
pixel_values: float32[B, 3, 384, 384], RGB, scaled to 0..1 then normalised with ImageNet mean[0.485, 0.456, 0.406]and std[0.229, 0.224, 0.225]. The image is resized (stretched) to 384x384, as in the upstreampredictcode. - Output
logits: float32[B, 300, 91], raw logits. Scores are per-class sigmoid values (the model is trained with sigmoid focal loss), not a softmax. - Output
pred_boxes: float32[B, 300, 4],cx, cy, w, hnormalised to 0..1 of the (stretched) input. There is no NMS: DETR style set prediction. - Pick vehicles by label string from
id2label.json(car,motorcycle,bus,truck), not by hard-coded index.
Use in BlurForge
BlurForge runs this model natively with ONNX Runtime (Android), threshold 0.35, two frames per second, inside an asset pack. Only vehicle boxes come from this model; plate geometry is derived from ML Kit text recognition on the vehicle crop.
Model tree for ipsilondev/rfdetr_nano_onnx
Base model
Roboflow/rf-detr-nano