3D Image Reconstruction from Multi-View Images using the Encoder-based Feature Map Generation

Abstract

The geometric feature of the image acquisition-based computer vision application utilizes the 3D image reconstruction for the required scene. The inaccurate reconstruction results and the enormous memory requirement limit the method’s performance. The deep learning-based methods provide promising results with elevated accuracy for computer vision applications. Hence, this research introduces a novel framework based on the deep learning strategy for the 3D image reconstruction of objects from multiple view images. The encoder-decoder-fusion-refiner module is proposed for the image reconstruction, in which the encoder module is proposed to generate the feature map. The significant attributes are taken out from the input multi-view image using the ResNet-50 and AlexNet pre-trained models, and then the channel and spatial information are gathered for obtaining the feature map. The most significant attribute section helps to reduce the computation complexity and elevates the accuracy of the reconstruction process. The newly devised method is evaluated by considering measures like loss and Intersection over Union (IoU) and obtaining the values of 1.04E+00, 1.11E+00, and 2.12E-01 for IoU_3D, IoU_2D and loss, respectively.