A Comparative Study of VoxelNet and PointNet for 3D Object Detection in Car by Using KITTI Benchmark

A Comparative Study of VoxelNet and PointNet for 3D Object Detection in Car by Using KITTI Benchmark

Harish S. Gujjar
DOI: 10.4018/IJICTHD.2018070103
(Individual Articles)
No Current Special Offers


In today's world, 2D object recognition is a normal course of study in research. 3D objection recognition is more in demand and important in the present scenario. 3D object recognition has gained importance in areas such as navigation of vehicles, robotic vision, HoME, virtual reality, etc. This work reveals the two important methods, Voxelnet and PointNet, useful in 3D object recognition. In case of NetPoint, the recognition is good when used with segmentation of point clouds which are in small-scale. Whereas, in case of Voxelnet, scans are used directly on raw points of clouds which are directly operated on patterns. The above conclusion is arrived on KITTI car detection. The KITTI uses detection by using bird's eye view. In this method of KITTI we compare two different methods called LiDAR and RGB-D. We arrive at a conclusion that pointNet is useful and has high performance when we are using small scenarios and Voxelnet is useful and has high performance when we are using large scenarios.
Article Preview

1. Introduction

A Point Net is based on cloud for detecting 3D objects which is a very burning concept in a real-world application which includes autonomous guiding of the object (Engelcke, Rao, Wang, Tong, & Posner; 2017; Krizhevsky, Sutskever, & Hinton, 2012), robots used in housekeeping (Oh & Watanabe, 2002) and virtual reality/augmented (Park, Lepetit, & Woo, 2008). In comparison with the detection based on images, LiDAR helps in correctly localizing the objects and shapes (Li, 2017) are characterized by giving valid depth information of an object. Although, objects which are not similar, LiDAR has scattered point clouds and has different point density, because of this reason of non-uniform division of the 3D space, sensors having an effective range, pose related and occlusion.

In many applications there is a need and urge to know the 3D method such as driving autonomously and augmented reality (AR). As there is more demand for sensors which detects 3D used in the mobiles and vehicles of autonomous, large amount of data is apprehended and processed. In this paper the author has studied the two methods of capturing 3D data and detection in the captured data. In the first method 3D sensors is used to record the data and classify the physical object category by boxes of 3D. 3D objet detection is a burning problem, the presentation of point cloud data and the uses of deep net architectures are important. Large number of works converts 3D point cloud to an image with the help of projection (Dorai & Jain, 1997; Girshick, 2015) or through quantization of grids by volumetric (Riegler, Ulusoys, & Geiger, 2016; Girshick, 2015) on the application of convolution networks.

A very short time ago Qi et al. (2017) PointNet was presented, actually an end-to-end intelligent neural network that follows the point wise property without deviation from the point clouds. This technique illustrates benchmarking results used on 3D recognition by dividing the object into parts and studying each part systematically.

In Qi, Yi, Su, and Guibas’s work (2017), an upgraded version of PointNet was unfolded which used to understand the existing object structures under different measures. To obtain more accurate results, these two methods have different characters on all the loaded data points (~1k points). Therefore, point clouds of the type acquire using LiDARs contain ~100k points, guiding the structures as in (Qi, Yi, Su, & Guibas, 2017; Qi, Su, Mo, & Guibas, 2017) outputs high calculation and storage space requirements by using region proposal network (RPN) (Ren, He, Girshick, & Sun, 2015).

Complete Article List

Search this Journal:
Volume 15: 1 Issue (2023): Forthcoming, Available for Pre-Order
Volume 14: 4 Issues (2022): 1 Released, 3 Forthcoming
Volume 13: 4 Issues (2021)
Volume 12: 4 Issues (2020)
Volume 11: 4 Issues (2019)
Volume 10: 4 Issues (2018)
Volume 9: 4 Issues (2017)
Volume 8: 4 Issues (2016)
Volume 7: 4 Issues (2015)
Volume 6: 4 Issues (2014)
Volume 5: 4 Issues (2013)
Volume 4: 4 Issues (2012)
Volume 3: 4 Issues (2011)
Volume 2: 4 Issues (2010)
Volume 1: 4 Issues (2009)
View Complete Journal Contents Listing