MVDUSt3R connects with InstantSplat

[wanghaonan-whn]

After MVDUSt3R infers and generates files in colmap format and then connects with InstantSplat, the Gaussian effect appears to be blurry. However, after upsampling the point cloud, the Gaussian effect improves but is still quite blurry. What could be the possible issue?

[recordmp3]

maybe camera pose is not accurate enough for test-time optimzation? Can you enabling the camera pose also to be learnable parameters?

We noticed that MV-dust3r is good at the correctness of the whole scene, but may have local inaccuracy. While dust3r with global optimization is good at pairwise finer accuracy, while sometimes suffering wrong matching, which will lead to whole scene failure.

Just brain storming, another heavier way to alleviate your problem may be a few iterations of dust3r+global optimization given connectivity from MV-dust3r, which will help removing the potential wrong matching, while keep everything not to be too slow.

[wanghaonan-whn]

Thank you for your answer. I have already tried modifying the pose and converting the relative pose into the pose in the same world coordinate system. The result has indeed improved, but it's still a bit blurry, and I feel it can be further optimized.