Wednesday, March 3, 2010

Towards Internet-scale Multi-view Stereo

Yasutaka Furukawa is back.

After PBA, PMVS et PMVS2. He and the team show very good result with large image collection.

On previous approach we notice a lot of parasite points (Easily visible on Photosynth => 3D points that results of bad correspondences...) but in this recent work they seems all filtered out. The only bad point is that some planar surface was not reconstructed ( windows, stained glass window). The part that are not reconstructed on the ground are due to detected moving people I think. So it's curious that from the number of images used the method is not capable of handling few shot to fill the remaining holes. On another side, the texture look clean and detailed !

It looks like the Daniel Martinec Round patches, but it's clearly more (See the thumbnail).

Source : Yasutaka Furukawa

No comments:

Post a Comment