RT-DETR r18vd (COCO), ONNX, patched for WebGPU
Person detector used by redact-web, combined with YuNet to find heads and faces to redact, in the browser.
- Upstream: https://huggingface.co/PekingU/rtdetr_r18vd, via the ONNX export at https://huggingface.co/onnx-community/rtdetr_r18vd.
- License: Apache-2.0, the same as upstream (copy in
LICENSE). - Training data: COCO.
Our modifications
Built by tools/models/rtdetr/export.py:
- Every
AveragePoolwithceil_mode=1is set toceil_mode=0, because onnxruntime-web's WebGPU backend rejectsceil_mode=1. At a 640x640 input every pooled map is even-sized, so ceil and floor give the same shapes; the export checks this and checks the outputs are unchanged up to float rounding. - No weights are changed.
Files
| File | Description |
|---|---|
onnx/model.onnx |
fp32 graph |
onnx/model_fp16.onnx |
fp16 graph with float32 inputs and outputs |
redact-web pins this repository at a commit and verifies every file's SHA-256 before use.