Estimation of volume of a solid object From Three Dimensional Point Clouds Generated By Convolutional Neural Networks based Semantic Segmentation

Volume 6 Issue 1 March - May 2019

Research Paper

Estimation of volume of a solid object From Three Dimensional Point Clouds Generated By Convolutional Neural Networks based Semantic Segmentation

Radhamadhab Dalai*

* Department of Computer Science & Engineering,Birla Institute of Technology, Ranchi, India.

Dalai, R. (2019). Estimation of volume of a solid object From Three Dimensional Point Clouds Generated By Convolutional Neural Networks based Semantic Segmentation.i-manager’s Journal on Pattern Recognition, 6(1), 27-34. https://doi.org/10.26634/jpr.6.1.16458

Abstract

Generating three dimensional point cloud for an object in image has found many applications in used in many computer vision systems. In this work a convolutional neural network based semantic segmentation has been used to find region of interest in an image. The region of interest has been represented as point clouds in three dimensional space. Then using image processing technique area based filter operations have been applied to find the total surface area. Finally adding all these small volumes total volume has been calculated. A large number of algorithms have been adapted reconstruction methods have been experimented and tested only for uniform backgrounds, which is disadvantageous for the applications on real images which consists of complex nonuniform regions. In this work semantic segmentation has been used to partition the regions into similar instance based regions. We have used UNET model for the region based segmentation. Then using encoderdecoder scheme the 3D point cloud has been generated after merging pixel clouds. This paper proposes an end-to-end efficient generation network, which is composed of an encoder, a 3D image model, and a decoder. First, a single-view image of object and a nearest-shape retrieval has been formed from UNET are fed into the network; then, the two encoders are merged adaptively according to their homo-graphic or similarity in nature. Then decoder generates fine-grained point clouds from the pixel clouds generated from multiple view images. Each point in the cloud represents a weight according the intensity and color information from which the density and volume of object has been calculated. The experiments on uniform background images show that our method attains accuracy 12 to 15 %margin compared with volumetric and point set generation methods particularly toward large solid objects, and it works multiple view angles as well.