Relaxing Accurate Initialization Constraint for 3D Gaussian Splatting
Keywords: Novel View Synthesis, 3D Gaussian Splatting
TL;DR: We train 3DGS from random point clouds.
Abstract: In this work, we investigate the limitations of the 3D Gaussian Splatting (3DGS) optimization scheme, revealing why it undergoes significant performance drops when initialized with noisy or random point clouds. Through in-depth analysis, we identify a key limitation of the 3DGS optimization: limited Gaussian transportability. Since Gaussians are optimized solely based on image photometric loss, the optimization tends to overfit the parameters of the projected Gaussians to improve reconstruction at their current positions, rather than relocating them to more optimal locations. This leads to producing under-reconstructed regions when starting with noisy or random initialization, failing to transport Gaussians to correct locations. Based on our findings, we propose RAIN-GS (Relaxing Accurate Initialization Constraint for 3D Gaussian Splatting), a set of simple yet effective modifications, including initializing sparse Gaussians with large variances, progressive Gaussian low-pass filtering, and an Adaptive Bound-Expanding split algorithm. These modifications enable Gaussians to effectively redistribute across the scene, capturing both coarse structure and fine details. By addressing the inherent limitations of 3DGS, RAIN-GS allows effective training even with random point clouds, significantly enhancing reconstruction quality.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 6403
Loading