AnnotateIt/rtdetr-r18vd-coco-onnx

RT-DETR R18vd COCO — verified ONNX mirror for AnnotateIt

This repository is an AnnotateIt-maintained, byte-identical mirror of the FP32 ONNX graph from onnx-community/rtdetr_r18vd. It is based on the official Apache-2.0 PekingU/rtdetr_r18vd checkpoint.

This is not an official PekingU or RT-DETR ONNX release. AnnotateIt did not alter or quantize the graph; the mirror adds an explicit license, immutable provenance, checksums, and independent numerical/runtime validation.

Artifact

File	Value
ONNX graph	model.onnx
Precision	FP32
Opset	16
Declared input	pixel_values: float32 [batch, channels, height, width]
Validated input	float32 [1, 3, 640, 640]
Outputs	logits: float32 [batch, 300, 80]; pred_boxes: float32 [batch, 300, 4]
Size	82,572,357 bytes
SHA-256	11843b02455cc24009aed24d4c40db721b1093be5ccd6bbe7b9c441abb1d0558

Preprocessing and outputs

1. Convert the image to RGB.
2. Resize to 640×640 with bilinear interpolation (stretch, without letterboxing).
3. Rescale channels to [0, 1] by dividing by 255; do not apply mean/std normalization.
4. Convert HWC to NCHW and add a batch dimension.
5. Apply sigmoid to logits and take the best COCO class score per query.
6. Interpret pred_boxes as normalized cx, cy, width, height and map them to the original image size.

The model predicts the 80 COCO classes and does not require NMS in the reference postprocessor.

Validation

• ONNX checker and ONNX Runtime CPU inference passed.
• The unchanged graph was compared with official PyTorch checkpoint revision ac77a11ff0170a41b771c03264987f8ce2b0d753 on a real COCO image.
• PyTorch and ONNX produced the same five detections at threshold 0.3, with matching classes, scores, and boxes.
• The graph ran successfully through AnnotateIt's onnxruntime-web 1.24.3 WASM runtime.

See validation-report.json for the measured differences and PROVENANCE.md for the pinned sources. Timings are smoke-test observations, not benchmarks.

License

RT-DETR and the official checkpoint are licensed under Apache License 2.0. The full text is included in LICENSE. Retain the license and attribution when redistributing the model.

Official project: lyuwenyu/RT-DETR