rfdetr-onnx

rfdetr-onnx

RF-DETR exported to ONNX for use as ffrwd model pins. The network is Roboflow's RF-DETR, Apache-2.0 for the Nano through Large sizes. This repo holds faithful exports of four checkpoints, nothing more: two of Roboflow's own, one third-party fine-tune for faces, and one fine-tune of our own for licence plates.

All are fp32, exported by the rfdetr pip package's own model.export(format="onnx") at each variant's own training size.

Hashes

The checkpoints each export started from. The two official ones are as the rfdetr package fetched them into ~/.roboflow/models; the face one is as downloaded from its repo; the plate one is the EMA weights our training run wrote, kept by hash rather than here.

The two COCO exports were made with rfdetr==1.10.0, the face and plate exports with rfdetr==1.10.1 (scripts/export_face.py, scripts/export_plate.py), each with torch from the package's own dependency set, opset 17, IR 8. The exporter does not simplify the graph: onnxsim is a listed dependency it never calls, so what is here is the graph as exported.

Model I/O

The input is planar RGB in NCHW, the whole frame stretched to the square with no aspect padding, scaled to 0..1 and then normalized with ImageNet's mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225). The graphs do not bake that in: their first node reads the input tensor directly.

dets is cx, cy, w, h, each a fraction of the picture. labels is raw class logits, so a confidence is sigmoid of the logit. For the two COCO exports the logits are in COCO's 91-slot category numbering - slot 1 is person, 62 chair, 72 tv - and a query's class is the largest of them. masks is raw mask logits per query over the whole square at quarter resolution; a pixel is inside its instance where the logit crosses zero. A DETR head returns no duplicates, so there is no NMS.

The face export

rf-detr-medium-face.onnx is RF-DETR Medium (DINOv2 backbone) as fine-tuned for single-class face detection by Herojayjay/RFDETR-Face-Detection, Apache-2.0, on a Kaggle face dataset of about 16,700 annotated images, at 576x576. It is not Roboflow's checkpoint and not trained on WIDER FACE; the WIDER FACE figures below are a check, not its benchmark.

Its head is one logit wide: labels is 1x300x1, index 0 is the face, and there is no background slot. The checkpoint's own class_embed is (1, 256), so the export loads it with num_classes left unset and lets the package align the head to the checkpoint. Passing num_classes=1 would tile the trained row into a second, invented class.

Measured on 200 WIDER FACE validation images (3,378 annotated faces) at a single 576 pass, matching at IoU 0.5, face size as the square root of the box area on the original image:

Small faces dominate that benchmark (2,649 of the 3,378 are under 32 px), and a single 576 pass cannot resolve them; the medium and large columns are the figures that describe a face that fills more of the frame.

The plate export

rf-detr-medium-plate.onnx is RF-DETR Medium fine-tuned by us for single-class licence plate detection, at 576x576, from Roboflow's rf-detr-medium.pth with rfdetr==1.10.1, on Open Images V7's "Vehicle registration plate" boxes (annotations CC BY 4.0, images CC BY 2.0) with the group-of boxes dropped: 4,497 training images carrying 6,554 plates, 724 validation images carrying 985. Twenty epochs, batch 8 with two steps of gradient accumulation, on one RTX 4090; validation mAP 50:95 peaked at 0.655 after the fourth epoch and the EMA weights of that epoch are what was exported (scripts/export_plate.py). The training data is what the model knows: photographs, mostly of parked and passing cars, plates from many countries, few of them at night.

Its head is two logits wide: rfdetr 1.10.1 lays a one-class head out as the class plus one spare slot, so labels is 1x300x2, index 0 is the plate, and index 1 was never trained toward anything and stays quiet - on the 2,065 test images below not one query's spare logit cleared 0.5. There is no background slot to subtract.

Measured on Open Images' own test split, 2,065 images holding 2,835 plates, at a single 576 pass, matching at IoU 0.5, plate size as the box's width on the original image:

Of the 2,835 plates, 580 are under 32 px wide, and a single 576 pass cannot resolve most of them; the two right-hand columns describe a plate that fills more of the frame.

Not here

RF-DETR XLarge and 2XLarge, and their segmentation counterparts, are under Roboflow's PML 1.0 rather than Apache-2.0, and are not exported here. The size names changed in 2026: the pre-rename "large" (487 MB fp32, 126M params) is today's XLarge, so check parameter counts and file sizes rather than names.