TetSphere Splatting: Representing High-Quality Geometry with Lagrangian Volumetric Meshes

Authors: Minghao Guo, Bohan Wang, Kaiming He, Wojciech Matusik

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on multi-view and single-view reconstruction highlight Tet Sphere splatting s superior mesh quality while maintaining competitive reconstruction accuracy compared to state-of-the-art methods. Additionally, Tet Sphere splatting demonstrates versatility by seamlessly integrating into generative modeling tasks, such as image-to-3D and text-to-3D generation. For evaluation, we conduct quantitative comparisons on two tasks to assess the geometry quality: multi-view and single-view reconstruction. In addition to the commonly used metrics for evaluating reconstruction accuracy, we introduce three metrics to evaluate mesh quality, focusing on key aspects of 3D model usability: surface triangles uniformity, manifoldness, and structural integrity. Compared to state-of-the-art methods, Tet Sphere splatting demonstrates superior mesh quality while maintaining competitive performance on other metrics.
Researcher Affiliation Academia Minghao Guo1 , Bohan Wang1,2 , Kaiming He1, Wojciech Matusik1 1MIT CSAIL, 2National University of Singapore
Pseudocode No The paper describes an algorithm for Tet Sphere initialization and provides its mathematical formulation in Appendix G, but it does not present structured pseudocode or an algorithm block.
Open Source Code Yes Code is available at https://github.com/gmh14/tssplat.
Open Datasets Yes For multi-view reconstruction, we follow Son et al. (2024) and use four closed-surface models from the Thingi32 dataset (Zhou & Jacobson, 2016), four open-surface models from the Deep Fashion3D dataset (Heming et al., 2020), and two additional models with both closed and open surfaces from the Objaverse dataset (Deitke et al., 2023). We also add the challenging sorter shape from Google Scanned Objects (GSO) dataset (Downs et al., 2022) as it features slender structures. For single-view reconstruction, following prior research (Liu et al., 2023c; Long et al., 2023), we use the GSO dataset for our evaluation, which covers a broad range of everyday objects.
Dataset Splits Yes For multi-view reconstruction, we follow Son et al. (2024) and use four closed-surface models from the Thingi32 dataset (Zhou & Jacobson, 2016), four open-surface models from the Deep Fashion3D dataset (Heming et al., 2020), and two additional models with both closed and open surfaces from the Objaverse dataset (Deitke et al., 2023). We also add the challenging sorter shape from Google Scanned Objects (GSO) dataset (Downs et al., 2022) as it features slender structures. For single-view reconstruction, following prior research (Liu et al., 2023c; Long et al., 2023), we use the GSO dataset for our evaluation, which covers a broad range of everyday objects. The evaluation dataset aligns with those used by Sync Dreamer and Wonder3D, featuring 30 diverse objects ranging from household items to animals.
Hardware Specification Yes We report the maximal batch size of 256 256 images that can occupy a 40GB A100 and the run-time speed for training with batch size 4.
Software Dependencies No Our implementation of the Tet Sphere initialization algorithm is developed in C++ and uses the Gurobi linear programming solver. The optimization of geometric energies is implemented using CUDA as a Py Torch extension to enhance computational efficiency.
Experiment Setup Yes In our experiments, we choose w1 = 5 10 6, w2 = 2 10 5 and apply a cosine increasing schedule.