Submitted by Martin Bokeloh.

Submission data

Full nameScanNet (re-implementation with flipped tsdf values)
DescriptionThe method follows the volumetric segmentation approach described in:
ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes
Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, Matthias Nießner

However, we apply the following changes:

1. Flipped TSDF
We use the flipped TSDF representation as described in:
Semantic Scene Completion from a Single Depth Image
Shuran Song Fisher Yu Andy Zeng Angel X. Chang Manolis Savva Thomas Funkhouser

2. We don't use free-space information.
The only input to the network is a 61x31x31 volume with flipped TSDF values.

3. We only use the mesh as input
The method computes a truncated distance field from the input mesh directly without using the provided depth images at all.
Publication authorsMartin Bokeloh, Juergen Sturm
Input Data TypesUses Geometry        Uses 3D
Programming language(s)C++/Python/Tensorflow
HardwareV100
Submission creation date4 Sep, 2018
Last edited5 Sep, 2018

3D semantic label results

Infoavg ioubathtubbedbookshelfcabinetchaircountercurtaindeskdoorfloorotherfurniturepicturerefrigeratorshower curtainsinksofatabletoiletwallwindow
0.3830.2970.4910.4320.3580.6120.2740.1160.4110.2650.9040.2290.0790.2500.1850.3200.5100.3850.5480.5970.394