SCISplat: 3D Gaussian Splatting from a Snapshot Compressive Image
Keywords: 3D Vision, Snapshot Compressive Imaging, Gaussian Splatting
Abstract: In this paper, we investigate the potential of Snapshot Compressive Imaging (SCI) for efficiently recovering 3D scenes from a single temporally compressed image. SCI offers a cost-effective approach using a series of 2D masks to compress video data into a single image captured by 2D imaging sensors. However, traditional SCI reconstruction methods face challenges with generalization and maintaining multi-view consistency. Recent advances have introduced Neural Radiance Fields (NeRF) to estimate 3D scenes from SCI images, but NeRF’s implicit representation struggles to capture fine details and support fast training and rendering. To address these issues, we propose SCISplat, a 3D Gaussian Splatting-based framework for decoding SCI images and achieving high-quality scene reconstruction from a single SCI image. First, we design an initialization protocol that robustly estimates the initial point cloud and camera poses from an SCI image, leveraging a learning-based Structure-from-Motion method. Second, we integrate the SCI image formation model into the 3D Gaussian training process and jointly optimize the Gaussians and camera poses to enhance reconstruction quality. Experiments demonstrate that SCISplat surpasses state-of-the-art methods, achieving a 2.3 dB improvement in reconstruction quality and a 10× faster training speed. Furthermore, results on real-world datasets show that our approach produces cleaner and sharper details, underscoring its practical value.
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: 6859
Loading