Multi-View Semantic Scene Completion from A Single Depth Image

Citation

Ching, Jiun Yen and Wong, Lai Kuan and Kung, Fabian Wai Lee (2022) Multi-View Semantic Scene Completion from A Single Depth Image. In: Postgraduate Colloquium December 2022, 1-15 December 2022, Multimedia University, Malaysia. (Unpublished)

[img] Text
CHING JIUN YEN-foe.pdf - Submitted Version
Restricted to Repository staff only

Download (280kB)

Abstract

Semantic scene completion aims to jointly predict the complete geometry and categories of objects in a scene, given sparse input from single observation. Previous works have used either depth only or together with 2D-to-3D projected RGB features learned by a 2D segmentation network, all stemming from a single view. We introduce a novel input, a modified flipped truncated signed distance to encode both visible and occluded surface information, simulating the benefits of having multiple views at virtually no extra memory cost. To train a network to produce this input, we derive a new ground truth (GT) from existing 3D instance data. Our experiments show that the novel input produce promising results.

Item Type: Conference or Workshop Item (Poster)
Uncontrolled Keywords: Segmentation network, Image segmentation
Subjects: T Technology > TA Engineering (General). Civil engineering (General) > TA1501-1820 Applied optics. Photonics
Divisions: Faculty of Engineering (FOE)
Depositing User: Ms Rosnani Abd Wahab
Date Deposited: 28 Dec 2022 01:32
Last Modified: 28 Dec 2022 01:32
URII: http://shdl.mmu.edu.my/id/eprint/11020

Downloads

Downloads per month over past year

View ItemEdit (login required)