Inverse Rendering for High-Genus 3D Surface Meshes from Multi-view Images with Persistent Homology Priors
Xiang Gao, Xinmu Wang, Yuanpeng Liu, Yue Wang, Junqi Huang, Wei Chen, Xianfeng Gu · 2026
Read Inverse Rendering for High-Genus 3D Surface Meshes from Multi-view Images with Persistent Homology Priors by Xiang Gao, Xinmu Wang, Yuanpeng Liu, Yue Wang, Junqi Huang, Wei Chen, Xianfeng Gu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-12155.txt.
Introduction
Xiang Gao, Xinmu Wang, Yuanpeng Liu, and their co-authors rebuild 3D objects from images with help from topology. Rebuilding is ill-posed, since shape, look, and topology can each be read in many ways from flat views. High-genus objects, with many holes, are the hard case.
They add persistent homology priors to inverse rendering. Those priors pin the hole count and shape class while views fit the rest. A firm topology prior can stop smooth but wrong rebuilds. The work targets meshes for design and scans. High genus shapes test the prior more than plain shapes.
Catalog record
| CREATOR | Xiang Gao, Xinmu Wang, Yuanpeng Liu, Yue Wang, Junqi Huang, Wei Chen, Xianfeng Gu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 16 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-12155
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Inverse Rendering for High-Genus 3D Surface Meshes from Multi-view Images with Persistent Homology Priors Xiang Gao, Xinmu Wang, Yuanpeng Liu, Yue Wang, Junqi Huang, Wei Chen, Xianfeng Gu Abstract Reconstructing 3D objects from images is inherently an ill-posed problem due to ambiguities in geometry, appearance, and topology. This paper introduces collaborative inverse rendering with persistent homology priors, a novel strategy that leverages topological constraints to resolve these ambiguities. By incorporating priors that capture critical features such as tunnel loops and handle loops, our approach directly addresses the difficulty of reconstructing high-genus surfaces. The collaboration between photometric consistency from multi-view images and homology-based guidance enables recovery of complex high-genus geometry while circumventing catastrophic failures such as collapsing tunnels or losing high-genus structure. Instead of neural networks, our method relies on gradient-based optimization within a mesh-based inverse rendering framework to highlight the role of topological priors. Experimental results show that incorporating persistent homology priors leads to lower Chamfer Distance (CD) and higher Volume IoU compared to state-of-the-art mesh-based methods, demonstrating improved geometric accuracy and robustness against topological failure. † † address: 1 Department of Computer Science, Stony Brook University, Stony Brook, NY, USA 2 Department of Applied Mathematics & Statistics, Stony Brook University, Stony Brook, NY, USA 3 Academy for Multidisciplinary Studies, Capital Normal University, Beijing, China {gao2, xinmuwang, yuanpliu, wang139, gu}@cs.stonybrook.edu, [email protected], [email protected] 1 Introduction Reconstructing high-genus 3D surface meshes from images arises in applications such as VR/AR, CAD, 3D printing, and scientific simulation or visualization, where preserving complex topological structures is crucial. While low-genus (genus-0) meshes can often be recovered from multi-view images, high-genus surfaces remain challenging due to topological ambiguity, which frequently causes tunnels and handles to collapse or disappear, leading to reconstructions with incorrect topology. Within a 3D surface mesh-based inverse rendering framework, the goal is to optimize the mesh so that its rendered appearance matches the input images. Existing studies [ 17 , 12 , 7 ] have attempted to reconstruct 3D surface meshes from multiview images using gradient-based optimization such as Adam [ 13 ] . Nicolet et al. [ 17 ] introduced an inverse rendering framework that enables accurate mesh reconstruction for genus-0 surfaces. Inspired by this framework, Jung et al. [ 12 ] extended it with a template-based approach and adaptive mesh density in feature-rich regions. Although these approaches achieve impressive 3D mesh reconstructions on low-genus surfaces, they fall short on complex high-genus structures, limiting their suitability for real-world applications. In addition, the standard practice of adopting uniform camera placement on a sphere is inherently biased and unjustified, as it assumes simple closed genus-0 geometries homeomorphic to a sphere. The rendered multiviews often produce vanishing or exploding gradients near tunnel regions, which appear as hole-like structures in the images. Moreover, severe occlusions across views can yield conflicting gradients, causing the optimization to compete across different viewpoints and ultimately leading to topological fragmentation in the reconstruction. Even when a 3D surface mesh is available, it is not straightforward to determine optimal camera placements, which further complicates robust reconstruction. This biased and unjustified strategy has been widely adopted in non-mesh-based deep generative image-to-3D studies [ 14 , 9 , 16 , 15 , 18 , 6 ] . Persistent Homology (PH) [ 4 ] , a powerful tool deeply rooted in algebraic topology, provides a principled way to detect topological invariants such as tunnel loops, offering a more explainable approach to place cameras. This paper proposes a collaborative inverse rendering strategy that incorporates persistent homology priors with a widely adopted uniform-sphere assumption to provide a justified basis for camera placements. The key insight is to place cameras near topological tunnels, which are topological invariants detected by persistent homology, thereby capturing regions prone to gradient vanishing, occlusion, or topological collapse. Compared to the state-of-the-art mesh-based inverse rendering method of Nicolet et al. [ 17 ] , our strategy shows stronger robustness against topological failure, achieving lower Chamfer Distance (CD) and higher Volume IoU. Figure 1: The pipeline of computing a persistent diagram. (a). Comparison between Čech Complex \mathcal{C}_{\varepsilon} and Vietoris-Rips Complex \mathcal{R}_{\varepsilon} [ 8 ] . (b). The evolution of filtered Cech complexes with varying scale parameters. (c). Persistent Diagram. 2 Methodology 2.1 Problem Formulation We formulate the inverse rendering problem for high genus surface meshes as: \displaystyle\arg\min_{\mathbf{x}} \displaystyle\Phi(R(\mathbf{x}))+\mathbf{w_{1}}\operatorname{tr}(\mathbf{x}^{T}\mathbf{L}\mathbf{x}) (1) \displaystyle\text{s.t.} \displaystyle\det(\mathbf{J}_{R}^{(k)})>0,\quad\forall k\in\{1,\dots,|F|\} where \mathbf{x}\in\mathbb{R}^{n\times 3} are mesh vertex positions, \mathbf{L}\in\mathbb{R}^{n\times n} the uniform bi-Laplacian [ 1 , 10 ] , R(\cdot) the rendering function, \Phi(\cdot) the rendering loss, \operatorname{tr}(\cdot) the matrix trace, |F| the number of faces, and \mathbf{J}_{R}^{(k)} the Jacobian of triangle k . Reformulating Equation 1 as an unconstrained optimization gives: \displaystyle\arg\min_{\mathbf{x}} \displaystyle\Phi(R(\mathbf{x}))+\mathbf{w_{1}}\operatorname{tr}(\mathbf{x}^{T}\mathbf{L}\mathbf{x}) (2) \displaystyle+\mathbf{w_{2}}\sum_{k=1}^{|F|}\left(\min\{0,\det(\mathbf{J}_{R}^{(k)})\}\right)^{2} which can be optimized by gradient descent, with \mathbf{w_{1}} enforcing smoothness and \mathbf{w_{2}} penalizing triangle inversion. 2.2 Preliminary on Persistent Homology (PH) Persistent homology is a central tool in Topological Data Analysis (TDA) for tracking the evolution of topological features across scales [ 5 , 3 ] . A simplicial complex \mathbb{K} is a collection of simplices closed under inclusion and intersection, where a k -simplex \sigma^{k}=[x_{0},\dots,x_{k}] is the convex hull of k+1 affinely independent points, such as vertices ( 0 -simplices), edges ( 1 -simplices), triangles ( 2 -simplices), and tetrahedra ( 3 -simplices). For a point set \{x_{\alpha}\}\subset\mathbb{R}^{n} , the Čech complex \mathcal{C}_{\varepsilon} includes a simplex if the corresponding closed \varepsilon/2 -balls intersect, while the Vietoris-Rips complex \mathcal{R}_{\varepsilon} includes a simplex if all pairwise distances between its vertices are at most \varepsilon . A filtration is then a nested sequence of complexes \emptyset=\mathbb{K}_{-1}\subset\mathbb{K}_{0}\subset\cdots\subset\mathbb{K}_{n}=\mathbb{K} , parameterized by \varepsilon , that captures topological changes across scales. In this setting, a k -chain is a linear combination of k -simplices forming the group C_{k}(\mathbb{K}) , with the boundary operator \partial_{k}:C_{k}\to C_{k-1} satisfying \partial_{k-1}\circ\partial_{k}=0 . Closed chains ( \partial_{k}\gamma=0 ) form the subgroup Z_{k} , exact chains ( \gamma=\partial_{k+1}\sigma ) form B_{k} , and the corresponding homology groups are defined as H_{k}(\mathbb{K})=Z_{k}(\mathbb{K})/B_{k}(\mathbb{K})=\ker\partial_{k}/\operatorname{im}\partial_{k+1} , distinguishing cycles from boundaries. The inclusion maps of a filtration induce homomorphisms between homology groups, 0=H_{p}(\mathbb{K}_{-1})\to H_{p}(\mathbb{K}_{0})\to\cdots\to H_{p}(\mathbb{K}_{n}) , thereby tracking the birth and death of features across scales. 2.3 Persistent Homology for Handle and Tunnel Loops We compute handle and tunnel loops of closed surfaces using persistent homology. Given a filtration of simplicial complexes \emptyset=\mathbb{K}_{-1}\subset\mathbb{K}_{0}\subset\cdots\subset\mathbb{K}_{n}=\mathbb{K} , each simplex is classified as positive if it creates a new homology class or negative if it destroys one. A positive simplex gives rise to a non-exact cycle, while a negative simplex pairs with the youngest active generator in its boundary, ending its persistence interval. For a genus- g closed surface M , the first homology group H_{1}(M) has rank 2g , so persistent homology yields 2g unpaired edges as fundamental loops; adding interior simplices pairs g of them with interior faces to form handle loops, while adding exterior simplices pairs the other g with exterior faces to form tunnel loops. The full procedure, including the Pair Algorithm for generator–killer matching, the Mark Loop Algorithm to trace cycles, and Null Homological Cycle Detection to discard null-homologous loops, followed by Birkhoff curve shortening for refinement, is illustrated in Fig. 1 , where the computed handle and tunnel loops are shown in Fig. 2 . Algorithm 1 Pair Algorithm 1: c\leftarrow\partial_{p}\sigma ; \tau\leftarrow