SuperPoint detection threshold and low matching speed
Closed this issue · 2 comments
In the case of setting SuperPoint detection_threshold to 0.05, LightGlue matching tooks lots of time sometimes. Why this can be happen? I guess it's due to extracted keypoint count is not equal to 512.
From your other issue, I'm assuming this happens when the TensorRT Execution Provider is used. The only explanation I can think of is that TensorRT encounters an unexpected shape and has to rebuild the engine/profile.
Are you using the TRT options like those passed here? If so, you can try to confirm that the number of keypoints returned by SuperPoint does indeed fall in the range between 1 and max_num_keypoints.
Thanks for your reply. I missed TRT options. I think it's better to use those options in the readme at master branch