Improved 3D Reconstruction for Images having Moving Object using Semantic Image Segmentation and Binary Masking

摘要

In computer vision, 3D reconstruction is a method for the measurement of figures and real substances. Structure from Motion (SFM) is a method for assessing three-dimensional structures from two-dimensional images. However, in 3D reconstruction because of the existence of moving object in the images, it creates some confusions like whether the object or the camera or both are moving. These confusions lead to inaccurate camera motion when the SFM algorithm reconstructs the trail of the camera with respect to a moving object. Moreover, problems like faulty camera positions and wrongly placed map objects occur because of 3D reconstruction with a moving object. This paper attempts to implement a way to resolve this problem for most moving and momentary objects. By using the semantic understanding of the images, this issue can be comprehended. Applying binary masking on the semantically segmented image we can classify the resulted image as excluded and included class which is considered to be black and white part of the binary masked image respectively. After 3D reconstruction of the binary masked image, it shows the significant non-moving portions of the scene and the camera motion with respect to it and solves the problem of 3D reconstruction for moving objects.