University of Twente Student Theses


Learned clustering for 3D object segmentation

Natarajan, Sriram (2020) Learned clustering for 3D object segmentation.

[img] PDF
Abstract:Applications related to autonomous driving, urban planning and asset monitoring rely on accurate information about the objects and their location in real world coordinates. Identifying stationary objects is one such application that finds importance in urban planning and asset monitoring, for instance: detection of roadside billboards, lamp posts etc. With the availability of point cloud representations of the environment, several approaches have been proposed for detection and segmentation of stationary objects in 3D. The detection of billboards is one such application which is challenging because of its incoherent visibility in multi-view images and absence of depth information due to its shape. This paper proposes Joint SPLATNet3D for semantic-instance segmentation of stationary objects in the scene. The proposed network performs two tasks: predicts a semantic label and generates an instance embedding for every 3D point. The multi-task loss function enables the network to jointly optimize the two tasks. This paper describes the dataset generation and feasibility study of semantic and instance segmentation for billboards. The paper gives a comparative analysis of Joint SPLATNet3D and MT-PNet for both the tasks. Preliminary experiments on semantic segmentation show that SPLATNet3D gives an average IoU of 75% in comparison with MT-PNet which gives an IoU of 46%. Experiments on joint training show that Joint SPLATNet3D gives an IoU of 68% in comparison with MT-PNet which gives an IoU of 48% for semantic segmentation. The results of instance segmentation for both the networks do not show good improvements for this dataset.
Item Type:Essay (Master)
Cyclomedia Technology, Zaltbommel, Netherlands
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Electrical Engineering MSc (60353)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page