Image matching for 3D reconstruction

Abstract

In this dissertation, we introduce LoGLUE, a novel method of image matching for 3D reconstruction that produces results for matching images with low texture. Compared with current image matching methods, we design a backbone with the Convolutional Neural Network (CNN) to extract both coarse- and fine-level features. This dissertation’s distinct contribution is the robust framework with a cross-attention and self-attention layer from Transformer that we proposed. We design the coarse-level module to do the coarse-level matching and then use an attention-based graph neural network to design the coarse-to-fine module. The results demonstrate that our framework performance is better than other existing methods when the inputs include a large patch of low-texture images. Reconstructing scenes with poor texture quality is now possible with the suggested LoGLUE framework.