What's the difference between the relased models?
Closed this issue · 1 comments
There are many models under relase_v0.1.3, what is the difference between them?
I thought the 512 in the name refers to the image resolution. But when I use v0.1.3 code to export 512 models via command
python export.py --img_size 512 --extractor_type superpoint --lightglue_path weights/superpoint_lightglue_512_end2end_my.onnx --dynamic --end2end.
The performance of the exported onnx model is different from the relased superpoint_512_lightglue_end2end.onnx.
I didn't find optimize.py so I assume it has nothing to do with model optimization? But then it's not clear what causes the different models obtained and what makes these models different? Thanks!
Hi @JingruiYu
The numbers in the name (512, 1024, 2048) correspond to the --max_num_keypoints argument, not --img_size. The models under v0.1.3 take any input resolution as they were exported with --dynamic.