Zhen Wang*1 Shijie Zhou*1 Jeong Joon Park2 Despoina Paschalidou2 Suya You3 Gordon Wetzstein2 Leonidas Guibas2 Achuta Kadambi1
University of California, Los Angeles1 Stanford University2 US Army Research Laboratory3
CVPR 2023, VancouverRethinking latent topologies for fast and detailed implicit 3D reconstructions. Recent work (POCO CVPR’22) has used latent encodings for each point to preserve 3D detail. We introduce ALTO, which can alternate between latent topologies like grid latents and point latents to speed up inference and recover more detail, like the 3D reconstruction of a thin lamp-post.
An overview of our method. Given input surface points, we obtain an implicit occupancy field with iterative alternation between features in the forms of points and 2D or 3D grids. Then we decode the occupancy values for query points with a learned attention-based interpolation from neighboring grids.
An illustration of our ALTO encoder.} (Left) As an example, we show the ALTO block instantiated by alternating between two latent topologies: point and triplanes via an ‘‘in-network’’ fashion, i.e. within each level of an hourglass framework U-Net. ‘Concatenate’ refers to concatenation of the ALTO block output triplane in the downsampling stage and the ALTO block input triplane in the corresponding upsampling stage. (Right) We expand on ALTO block to illustrate the sequential grid-to-point and point-to-grid conversion. There are skip connections for both point and grid features between two consecutive levels in the ALTO U-Net.
Object-level comparisons on ShapeNet. On the car, ALTO recovers the detail of having both side mirrors.
Cross-dataset evaluation of ALTO and baselines by training on Synthetic Rooms and testing on real-world ScanNet-v2. Note the large conference-room table is missing in ConvONet (purple inset). The ladder (yellow inset) is a high-frequency surface and we believe our method is qualitatively closest.
Zhen Wang
Electrical and Computer Engineering Department
zhenwang@ucla.edu
Shijie Zhou
Electrical and Computer Engineering Department
shijiezhou@ucla.edu