generate high-quality point clouds

[wwj-53]

Do you plan to use it to generate high-quality point clouds in the future? I think using it to generate high-quality point clouds would have progressive significance for many tasks such as SLAM and 3D reconstruction.

[wyf2020]

Great question! In fact, our another CVPR 2024 work, DetectorFreeSfM (which ranked 1st in the Image Matching Challenge 2023), is a sfm framework specifically designed for detector-free matching. It can generate high-quality point clouds and complete high-quality reconstruction or triangulation. The code has been open-sourced and is ready for use.

[wwj-53]

I took a brief look and DetectorFreeSfM is based on LoFTR. Would the performance be better if we switch to Efficient LoFTR? Do you have any plans to use Efficient LoFTR to create point clouds?

[wyf2020]

Using Efficient LoFTR in DetectorFreeSfM is straightforward. You can check the file for examples of using LoFTR, MatchFormer, and AspanFormer, and apply the same principles to use EfficientLoFTR.