DUSt3R does not perform well when the view angle difference is too large in object scenario

[Sansju]

Congratulations for your great work!
Your paper says that the geometry-aware feature alignment module can distill geometric priors from dense stereo models during training. And for object scenario, your work is trained on Objaverse dataset. I have expilicitly experimented DUSt3R on Objaverse dataset and have some trouble. I worried about distilling this wrong geometry prior will get bad result. Here are two questions about object scenario:
1.Geometry-aware feature alignment loss is MSE between Ft and the concatenated point maps produced by DUSt3R given the t-th view It and the anchor view I1. But when generating novel views for synthetic 3D objects, view angle between It and I1 will be very large, and DUSt3R can't deal with it(in our experiment). How you overcome this trouble?
2.In our experiment DUSt3R does not perform well on some subset in Objaverse dataset, such as objects with low texture. Have you filtered the objects that the DUSt3R doesn't handle well? If so, how did you do it?
Looking for your apply.

[lg-li]

Thanks for your question!
Indeed, DUSt3R does not perform well with rendered textureless images and views with extremely large angle gaps. During our mixed training, we disabled the geometric feature alignment for samples from Objaverse that include textureless rendered images and target views on an orbit to avoid interference from DUSt3R's limitations.
This adjustment does not significantly impact our testing on Objaverse, as we have trained on a mixed dataset that includes another object-centric dataset CO3D.