Chin, Wai Lok (2021) Object Localization In 3D Point Cloud. Final Year Project, UTAR.
Abstract
Object localization in point clouds can help search for the target objects in the extensive 3D search space. It allows the post-operation of object recognition to operate on the objects more efficiently. There are many published works for object localization in 3D point clouds. Each approach has a unique architecture in its work. Thus, the frameworks used are not standardized like with 2D object localization frameworks. This work focuses on developing a method to locate objects in a point cloud and measure the objects’ three primary dimensions accurately. The intra and inter-comparison and evaluation of the selected work are conducted to discuss its significance in 3D object localization. Comparison and evaluation of method(s) are standardized by average precision outputted using the same evaluation metrics, the KITTI offline evaluation dataset. Point-GNN is selected as the approach for 3D object localization. It works best when iterated twice in the edges and vertices’ feature aggregation. Besides, Point-GNN scored second among the twelve 3D object localization approaches discussed. It achieves the AP predicted on the KITTI test 3D detection benchmark of 88.33 % for ‘easy’ car, 79.47 % for ‘moderate’ cars, and 72.29 % for ‘hard’ cars.
Actions (login required)
![[img]](http://eprints.utar.edu.my/style/images/fileicons/application_pdf.png)
