GaussianMatch: Adaptive Learning Continuous Surfaces for Light Field Depth Estimation
Loading...
Date
2025
Journal Title
Journal ISSN
Volume Title
Publisher
The Eurographics Association
Abstract
Light field (LF) depth estimation plays a vital role in computational imaging by reconstructing 3D structures from multiple viewpoints. However, images are merely discrete expressions of scenes due to the resolution constraints of cameras, leading to depth discontinuities and outliers-particularly in textureless or occluded regions, degrading reconstruction coherence. To address the challenges mentioned above, we propose GaussianMatch, a probabilistic depth estimation framework that models per-pixel depth as a learnable Gaussian distribution in continuous space. This scheme effectively alleviates the discretization problem of LF images by adaptively reconstructing continuous surfaces, while enabling uncertainty-aware optimization.Furthermore, the framework naturally fuses information among adjacent pixels and adapts each Gaussian's variance according to scene complexity, achieving robustness in both texture-rich and ambiguous regions. We further design GaussianNet, which regresses per-pixel Gaussian parameters and generates the final depth map via Gaussian accumulation. Extensive experiments on multiple LF benchmarks demonstrate that GaussianNet achieves state-of-the-art accuracy, with significant improvements in handling depth discontinuities and occlusions.
Description
CCS Concepts: Computing methodologies → Matching; Epipolar geometry; Computational photography; Image representations
@inproceedings{10.2312:pg.20251283,
booktitle = {Pacific Graphics Conference Papers, Posters, and Demos},
editor = {Christie, Marc and Han, Ping-Hsuan and Lin, Shih-Syun and Pietroni, Nico and Schneider, Teseo and Tsai, Hsin-Ruey and Wang, Yu-Shuen and Zhang, Eugene},
title = {{GaussianMatch: Adaptive Learning Continuous Surfaces for Light Field Depth Estimation}},
author = {Sun, Zexin and Chen, Rongshan and Wang, Yu and Cui, Zhenglong and Yang, Da and Li, Siyang and Huang, Xuefei and Sheng, Hao},
year = {2025},
publisher = {The Eurographics Association},
ISBN = {978-3-03868-295-0},
DOI = {10.2312/pg.20251283}
}