From NeRF to 3DGS: A Leap in Stereo Dataset Quality?
Keywords: 3D Gaussian Splatting (3DGS), Neural Radiance Fields (NeRF), Stereo Dataset Generation, Disparity Estimation, Deep Stereo Matching, Synthetic Data Quality
TL;DR: The paper presents a study on enhancing stereo dataset quality and disparity estimation accuracy by integrating 3D Gaussian Splatting (3DGS) in the place of Neural Radiance Fields (NeRF)
Abstract: Recent advancements in stereo matching, driven by deep learning techniques, have increased the need for datasets containing dense ground truth disparity labels. Yet, the rarity of real-world datasets with these labels presents significant challenges stemming from the difficulties in creating accurate dense depth maps. This often involves complex structured light setups, producing a constrained quantity of high-quality samples, or employing laser-based distance sensors, which offer more accessible but sparsely labelled and less accurate data. A promising development in this context is the utilization of Neural Radiance Fields (NeRFs), which leverage a minimal set of RGB images to synthesize stereo images with highly accurate dense disparity maps. Despite the high quality of synthesized images, NeRF-generated disparity maps exhibit a significant number of outliers, necessitating complex training paradigms for effective use. Our study investigates using 3DGS over NeRFs to produce stereo training views and dense disparity labels. We demonstrate that 3DGS offers enhanced accuracy in generating disparity labels and propose an efficient strategy for identifying and removing outliers, thereby significantly improving the disparity labels quality.
Submission Number: 56
Loading