Monocular 3D Reconstruction in Poorly Visible Environments

Abstract

3D reconstruction of real physical environments can be a challenging task, often requiring depth cameras such as LIDAR or RGB-D to capture the necessary depth information. However, this method is resource-intensive and expensive. To counter this problem, monocular 3D reconstruction has emerged as a research area of interest, leveraging deep learning techniques to reconstruct 3D environments using only sequences of RGB images, thus reducing the need for specialized hardware. Existing research has primarily focused on environments with good lighting conditions, leaving a gap in research for environments with poor visibility. In response, we propose a solution that addresses this limitation by enhancing the visibility of images taken in poorly visible environments. These enhanced images are then used for 3D reconstruction, resulting in the extraction of more features and producing a 3D mesh with improved visibility. Our solution employs a Generative Adversarial Network (GAN) to enhance the images, providing a complete pipeline from inputting images with poor visibility to generating an output mesh file for 3D reconstruction. Through visualization of these mesh files, we observe that our solution improves the lighting conditions of the environment, resulting in a more detailed and readable 3D reconstruction.