LAST: Latent Structure guided Gaussian Splatting from Monocular Human Videos

Sihao Liu, Shengyu Zhang, Sida Peng, chengfei lv, Fei Wu

Keywords: Digital Human, 3D reconstruction, Causal Inference

TL;DR: Disentangle meaningful latent factors and realistic dependencies from the input video frames, which allows for dynamic adjustments to the density and attributes of Gaussian points during the optimization process.

Abstract: Multiocular human reconstruction aims to create a high-quality 3D human representation from sparse video data. Recently, 3D Gaussian Splatting (3DGS) has demonstrated impressive results in multiocular human reconstruction tasks, exhibiting remarkable speed and accuracy. However, it encounters challenges in scenarios involving intricate clothing and dynamic postures. This problem may stem from pixel-level supervision during the 3DGS optimization process, which probably lead to spurious associations between unrelated visual features (e.g., misinterpreting clothing wrinkles as dependent on body occlusions rather than recognizing that both are influenced by complex postures). To address this issue, we propose the LAST framework for realistic 3D human reconstruction, which integrates a pre-trained Image-to-Point (I2P) model to enhance the 3D Gaussian Splatting optimization pipeline. The core of the LAST is to disentangle meaningful latent factors and realistic dependencies from the input video frames, which allows for dynamic adjustments to the density and attributes of Gaussian points during the optimization process. Experimental results demonstrate that our method significantly improves accuracy and realism in 3D human reconstruction compared to existing techniques, particularly in challenging scenarios involving complex posture and intricate clothing details.

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: 9238