The authors first examined how visual geometry transformers behave as image resolution and view count increase.
Fixed-size basis tokens make high-resolution 3D reconstruction faster and sharper
ReDT performs geometric reasoning in a compact representation whose size does not grow with image resolution, then projects the results back for dense reconstruction.
Chinese Tech
Abteen Arab · Guile Wu · Dongfeng Bai
Huawei Noah’s Ark Lab · University of British Columbia
Research Digest··3 min read
Arab, Wu and Bai introduce ReDT, a feed-forward transformer for reconstructing 3D scenes from multiple high-resolution images.
§