Submitted by Chris Choy.

Submission data

Full nameMinkowskiNet
DescriptionWe used the 3D variant of the 4D MinkowskiNet34, which consists of 42 (34 + 8) convolution layers and was trained for 120k iterations.

Next, we train the network from LR=1e-2 on both train+val set for another 120k iterations.

For implementation, we used the Minkowski Engine (https://github.com/StanfordVL/MinkowskiEngine).
Publication title4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks
Publication authorsC. Choy, J. Gwak, S. Savarese
Publication venueCVPR 2019
Publication URLhttps://arxiv.org/abs/1904.08755
Input Data TypesUses Color        Uses 3D
Programming language(s)python/pytorch/C++/CUDA
HardwareTitanRTX
Websitehttps://github.com/chrischoy/SpatioTemporalSegmentation
Source code or download URLhttps://github.com/chrischoy/SpatioTemporalSegmentation
Submission creation date27 Oct, 2018
Last edited1 Oct, 2019

3D semantic label results

Infoavg ioubathtubbedbookshelfcabinetchaircountercurtaindeskdoorfloorotherfurniturepicturerefrigeratorshower curtainsinksofatabletoiletwallwindow
permissive0.7360.8590.8180.8320.7090.8400.5210.8530.6600.6430.9510.5440.2860.7310.8930.6750.7720.6830.8740.8520.727