Submitted by Chris Choy.

Submission data

Full nameMinkowskiNet34
DescriptionWe used the 3D variant of the MinkowskiNet34, which consists of 42 (34 + 8) convolution layers.

We used the Minkowski Engine (https://github.com/chrischoy/MinkowskiEngine) to implement the network and it took ~1s to process an entire room in one go. The method will be made available to the public.
Input Data TypesUses Color        Uses 2D
Programming language(s)python/pytorch/C/C++/CUDA
HardwareGTX1080Ti
Submission creation date27 Oct, 2018
Last edited20 Nov, 2018

3D semantic label results

Infoavg ioubathtubbedbookshelfcabinetchaircountercurtaindeskdoorfloorotherfurniturepicturerefrigeratorshower curtainsinksofatabletoiletwallwindow
0.6790.8110.7340.7390.6410.8040.4130.7590.6960.5450.9380.5180.1410.6230.7570.6800.7230.6840.8960.8210.651