besit/rfdetr-onnx

RF-DETR Segmentation — ONNX

ONNX exports of the COCO-pretrained RF-DETR segmentation models by Roboflow, for use with the ONNX Segment node for Houdini Copernicus (downloaded automatically by the node).

The weights are Roboflow's, unmodified. They were only exported to ONNX with rfdetr 1.11.0 (model.export(format="onnx"), opset 17, static batch 1). RF-DETR is released under the Apache License 2.0, Copyright 2025 Roboflow, Inc.

File	Input	Queries	Mask	SHA-256
rfdetr-seg-nano.onnx	312×312	100	78×78	482f13bb698916dfcafbd98b2e563353721eb1d56f5d4851c6ada7bc9fa0ba94
rfdetr-seg-small.onnx	384×384	100	96×96	6e9ac79e34143c8cbd63b5e1c653c3108ea944ab08153433c761f72d7fef641c
rfdetr-seg-medium.onnx	432×432	200	108×108	33f97a068ea7de3892e4cbc3825e15c5418be818506d7f39f44e1fb1f7ab6abb
rfdetr-seg-large.onnx	504×504	200	126×126	c0aec7057051b13aced2592146fcdd92a94237c801186de4b5b28efc623adf06

Inputs and outputs

• input — [1, 3, S, S] float32, RGB in 0–1, normalized with ImageNet mean [0.485, 0.456, 0.406] and std [0.229, 0.224, 0.225], plain resize (no letterbox).
• dets — [1, Q, 4] boxes as normalized cx, cy, w, h.
• labels — [1, Q, 91] logits; apply a sigmoid. The index is the sparse COCO category id (1 = person).
• masks — [1, Q, S/4, S/4] mask logits for the whole image; a pixel belongs to the object where the logit is above 0 (probability 0.5).

Export and reference scripts: github.com/besit/yolo_onnx (rfdetr/).