v1.0.0: DeDoDe-ONNX-TensorRT

DeDoDe ONNX Models

This release provides exported ONNX models of the DeDoDe models. Both standalone models and end-to-end pipelines (*_end2end.onnx) are provided.

Individual Models

• detector_L_{N}.onnx: DeDoDe detector (L variant) with num_keypoints=N.
• descriptor_B_{N}.onnx: DeDoDe descriptor (B variant) with num_keypoints=N.
• matcher.onnx: Dual Softmax matcher.
• *_fp16.onnx: Half-precision variants.

End-to-end Models

• dedode_end2end_{N}.onnx: End-to-end DeDoDe pipeline with num_keypoints=N.
• *_fp16.onnx: Half-precision variants.

TensorRT-compatible DeDoDe ONNX Models

These models have undergone shape inference and can be executed on ONNXRuntime's TensorRT Execution Provider. Only end-to-end versions are provided as they work best with TensorRT. Note that TensorRT has a limitation which restricts the maximum number of keypoints to 3840.

• dedode_end2end_{N}_trt.onnx: TensorRT-compatible end-to-end DeDoDe pipeline with num_keypoints=N.

Environment Compatibility & Reproducibility

These models were exported and tested with the following framework versions:

• CUDA==11.8
• cuDNN==8.9.4
• TensorRT==8.6.1
• torch==2.0.1
• torchvision==0.15.2
• onnx==1.14.0
• onnxconverter-common==1.13.0
• onnxruntime-gpu==1.15.1