3D Reconstruction of Low-Resource Indoor Environments Using Deep Learning-Based Image Enhancement: A Case Study of 3D Gaussian Splatting

This study investigates whether deep learning–based image enhancement can improve the applicability of 3D Gaussian Splatting for three-dimensional scene reconstruction in low-resource indoor environments. 3D Gaussian Splatting provides high-quality 3D reconstructions when high-resolution, multi-view images are available; however, many indoor monitoring environments, including those in developing countries, mainly rely on low-cost CCTV and IP cameras, which suffer from various forms of image degradation such as blur, lossy compression, and reduced resolution. Because high-quality reference images of the same scene are often unavailable in real low-quality video, we first acquired highquality original images by capturing typical indoor scenes from multiple viewpoints using a fixed camera and a mobile device, and then deliberately generated a set of degraded images of the same scenes by applying Gaussian blur, lossy compression, and cropping at different levels. Using a pre-trained deep learning–based image enhancement model, we restored the degraded images to construct three conditions for each scene: high-quality original, degraded, and degradedplus-enhanced images, and applied the 3D Gaussian Splatting pipeline to attempt 3D reconstruction under each condition. The evaluation focused solely on whether reconstruction succeeded or failed for each condition, and compared the corresponding success rates. The results suggest that, even in developing countries or regions with poor infrastructure, it may be possible to improve indoor 3D modeling and safety monitoring by combining software-based image enhancement with 3D Gaussian Splatting while retaining existing low-resolution camera infrastructure, highlighting the potential of this approach as an appropriate technology.