Export script or Static Shape version for TensorRT?
Hello! First of all, thank you for sharing these models, they are extremely useful.
I am currently playiing with them into a C++ pipeline using ONNX Runtime with the TensorRT execution provider. However, I am facing performance issues because the current LightGlue Matcher uses dynamic input shapes (varying number of keypoints). This prevents TensorRT from building an optimized engine, causing it to fall back to standard CUDA with significant overhead due to memory copies.
Would it be possible to share the export script used to generate these files? I would like to re-export the matcher with fixed/static input dimensions (e.g., padded to exactly 2000 keypoints) to ensure TensorRT compatibility.
Alternatively, if you could upload a version with static shapes, that would be amazing as well.
Thanks in advance!