The combination of RaCo with SuperPoint and ALike does not yield good results.

[QIANXUNZDL123]

This is an excellent piece of work. However, during my usage, I found that RaCo must be paired with ALIKED to achieve good results. The problem is that ALIKED contains DCN (Deformable Convolution) operators, which cannot be exported to ONNX. I tried pairing RaCo with SuperPoint and ALike instead, but the results were extremely poor. I would like to ask: does RaCo necessarily need to be paired with ALIKED to achieve good results?

[sarlinpe]

Thank you for your interest!

I tried pairing RaCo with SuperPoint and ALike instead, but the results were extremely poor.

This is surprising. RaCo is not trained for a specific descriptor so it should work ok with SuperPoint, whose descriptors are indeed worse than ALIKED but not that much.

The problem is that ALIKED contains DCN (Deformable Convolution) operators, which cannot be exported to ONNX.

FWIW ALIKED can be successfully exported to ONNX: https://github.com/colmap/ALIKED-ONNX/tree/user/jsch/onnx-export/export It will be part of COLMAP's next release.

[QIANXUNZDL123]

感谢你的关注！

我尝试将RaCo与SuperPoint和ALike搭配使用，但效果非常差。

这很令人惊讶。RaCo没有针对特定描述符进行训练，所以应该能和SuperPoint兼容，SuperPoint的描述确实比ALIKED差，但也没那么严重。

问题在于ALIKED包含DCN（可变形卷积）算符，无法导出到ONNX。

仅供参考，ALIKED 可以成功导出到 ONNX：https://github.com/colmap/ALIKED-ONNX/tree/user/jsch/onnx-export/export 它将成为 COLMAP 下一个版本的一部分。

Thank you for taking the time to reply to me. I will export the ALIKED model based on the link you provided. I used SuperPoint with RaCo for matching but the results were indeed poor. Below are the result images: