Interactive Segmentation of RGB-D Indoor Scenes using Deep Learning

Rüthlein M, Köferl F, Mehringer W, Eskofier B (2020)


Publication Type: Conference contribution, other

Publication year: 2020

Event location: Virtual Conference

Abstract

Human segmentation of point clouds for the creation of datasets for deep learning is a tedious and especially time-consuming task. Interactive segmentation methods from the domain of RGB images reduce this time effort by using an iterative scheme of deep neural networks and human labelers. We apply this interactive segmentation scheme to the point cloud domain, using PointNet as our backbone and exploiting user marker-information by label propagation. The evaluation was based on a user study and suggests a significant increase in segmentation speed, which is traded off for mask quality. On average, our approach decreased segmentation time by 21.1%. We consider this work as an important first step into the domain of interactive point cloud segmentation.

Authors with CRIS profile

Additional Organisation(s)

How to cite

APA:

Rüthlein, M., Köferl, F., Mehringer, W., & Eskofier, B. (2020). Interactive Segmentation of RGB-D Indoor Scenes using Deep Learning. In Proceedings of the International Conference on Machine Learning; 2nd ICML 2020 Workshop on Human in the Loop Learning. Virtual Conference.

MLA:

Rüthlein, Maximilian, et al. "Interactive Segmentation of RGB-D Indoor Scenes using Deep Learning." Proceedings of the International Conference on Machine Learning; 2nd ICML 2020 Workshop on Human in the Loop Learning, Virtual Conference 2020.

BibTeX: Download