Nice point maps. Yet, deceiving point map reprojection consistency.

[YanNoun]

Hi,

I've following MapAnything with great attention and I've been able to fine-tune for my own scenarios. I've also pulled-in others model, namely Mast3r, for comparison. While I was initially pleased with MapAnything results, especially on point cloud completeness and consistency across various condition and scenes scales, I have yet been quite deceived with the reprojection consistency (backprojecting a pointmap in a another view) of MapAnything, especially compared to Mast3r.

While I understand that MapAnything hasn't been trained with 2D matching task in mind - and Mast3r has - I'm puzzled by how the final pointcloud looks great, but the pointmaps are not so accurately consistent wrt. the views.

I'd like to improve MapAnything on that front, and was wondering if you had any success or failure with using cross-view pointmap 3D consistency, or even using raymaps and relative poses to impose epipolar constraints across views ? Or is cross-views constraint a dead-end (and that's why MapAnything works well, view-wise constraints are the best, the data does the rest) ?

Thank you for your answr,

Cheers,

Yann