Unofficial ONNX conversion of Roboflow RF-DETR โ€” small

Single-file ONNX exporting the raw model outputs (no NMS, no sigmoid, no top-k, no resizing โ€” all pre/postprocessing stays outside the graph). Unofficial conversion; not endorsed by the upstream authors. Upstream-reported COCO AP: 53.0 (not re-measured here).

Class table is UNVERIFIED. RF-DETR emits a 91-slot DETR-family COCO layout (index 0 = background). Confirm the column count and no-object slot against this exact file (Netron + one known image) before relying on the labels.

Input contract

name input
dtype / layout float32, NCHW
shape [1, 3, 512, 512] (static)

Preprocessing: RGB โ†’ resize (stretch) to 512ร—512 โ†’ float32 รท255 โ†’ ImageNet mean/std after รท255 โ†’ HWCโ†’CHW โ†’ batch dim.

Output contract

output shape meaning
labels [1, 300, 91] raw class logits (apply sigmoid yourself)
dets [1, 300, 4] normalized cxcywh

91-class COCO (torchvision 91-slot), 300 queries, DETR-style set prediction โ€” no NMS.

Artifact

model.onnx SHA-256 850b7a24b573064f98f41605e4e400421420edf9b377e1704fa2079589cc3479 (114,439,180 bytes)

License

Apache-2.0 (see LICENSE / NOTICE). Upstream: Roboflow RF-DETR.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support