Submitted by Dario Rethage.

Submission data

Full nameFully-Convolutional Point Networks
DescriptionHybrid (point-to-voxel) network for deep learning on large-scale 3D data. Fully geometric - no color used.
Publication titleFully-Convolutional Point Networks for Large-Scale Point Clouds
Publication authorsDario Rethage, Johanna Wald, J├╝rgen Sturm, Nassir Navab, Federico Tombari
Publication venueECCV 2018
Publication URLhttps://arxiv.org/abs/1808.06840
Input Data TypesUses Geometry        Uses 3D
Programming language(s)Python, Tensorflow, Cuda
HardwareNvidia Titan X
Websitehttp://github.com/drethage/fully-convolutional-point-network
Source code or download URLhttps://github.com/drethage/fully-convolutional-point-network
Submission creation date18 Feb, 2019
Last edited3 Mar, 2019

3D semantic label results

Infoavg ioubathtubbedbookshelfcabinetchaircountercurtaindeskdoorfloorotherfurniturepicturerefrigeratorshower curtainsinksofatabletoiletwallwindow
permissive0.4470.6790.6040.5780.3800.6820.2910.1060.4830.2580.9200.2580.0250.2310.3250.4800.5600.4630.7250.6660.231