Thread by @ilkedemir

For the night session attendees of #CVPR2020, here is my completely non-scientifically chosen top five papers @cvpr for the second day. Visit them in the next 8 hours! (1/6)

1. DualConvMesh-Net: Joint Geodesic and Euclidean Convolutions on 3D Meshes

Authors: Jonas Schult, Francis Engelmann, Theodora Kontogianni, Bastian Leibe

Conv->(euclidean+geodesic) convs
Pooling->mesh simplification
6% mIoU increase and a nice paper!

#CVPR2020

2. Unsupervised Learning of Intrinsic Structural Representation Points

Authors: Nenglun Chen, Lingjie Liu, Zhiming Cui, Runnan Chen, Duygu Ceylan, Changhe Tu, Wenping Wang

Like categorical SIFT points in point clouds, for matching, recon, interpolation and more!

#CVPR2020

3. Deep 3D Capture: Geometry and Reflectance From Sparse Multi-View Images

Authors: Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, Ravi Ramamoorthi

You need only 6 images to capture both the geometry and the SVBRDF of an object! Hello (shiny) kitty :)

#CVPR2020

4. Local Deep Implicit Functions for 3D Shape

Authors: Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, Thomas Funkhouser

The SIF ellipsoids of last year become DIF this year, containing local deep features and increasing reconstruction capability.

#CVPR2020

5. StructEdit: Learning Structural Shape Variations

Authors: Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy J. Mitra, Leonidas J. Guibas

Let's extract the edits from a parse tree of a boxelized object with a VAE. Only box parameters and +/- nodes.

#CVPR2020

Bonus: Another shameless plug. If you want to not only edit the structure, but also create plausible geometry between nodes, check this shallow paper. It also comes with explicit patterns and a point cloud enhancing consensus model.

cv-foundation.org/openaccess/con…

Non-CVPR2020 :)