DCUDF2: Improving Efficiency and Accuracy in Extracting Zero Level Sets from Unsigned Distance Fields

TL;DR

DCUDF2 addresses the challenge of extracting accurate zero level sets from unsigned distance fields by combining an accuracy-aware loss with self-adaptive weights, a topology-correcting mechanism, activation masks, and optimization-direction safeguards. It builds on the DCUDF framework by reducing over-smoothing, lowering dependency on the iso-value $r$, and boosting runtime efficiency, validated through extensive experiments across noisy, complex, and multi-view UDFs. The results show improved geometric fidelity and topological accuracy over state-of-the-art methods, including complete elimination of non-manifold configurations in many cases. The approach enables robust, high-quality surface extraction from diverse UDFs and offers practical benefits for downstream applications requiring reliable manifold meshes.

Abstract

Unsigned distance fields (UDFs) allow for the representation of models with complex topologies, but extracting accurate zero level sets from these fields poses significant challenges, particularly in preserving topological accuracy and capturing fine geometric details. To overcome these issues, we introduce DCUDF2, an enhancement over DCUDF--the current state-of-the-art method--for extracting zero level sets from UDFs. Our approach utilizes an accuracy-aware loss function, enhanced with self-adaptive weights, to improve geometric quality significantly. We also propose a topology correction strategy that reduces the dependence on hyper-parameter, increasing the robustness of our method. Furthermore, we develop new operations leveraging self-adaptive weights to boost runtime efficiency. Extensive experiments on surface extraction across diverse datasets demonstrate that DCUDF2 outperforms DCUDF and existing methods in both geometric fidelity and topological accuracy. We will make the source code publicly available.

Figures (12)

• Figure 1: DCUDF, a leading method for extracting zero level sets from UDFs, often suffers from over-smoothing that compromises geometric details. Our enhanced approach improves upon DCUDF by more effectively preserving geometric fidelity and topological accuracy, while also enhancing runtime performance. It outperforms existing methods in maintaining detailed geometry and accurate topology.
• Figure 2: Conceptual illustration of the over-smoothing side effect in DCUDF, particularly pronounced when the target surface has rich geometric details and a large iso-value $r$ is used. In this illustration, $S$ represents the target zero level set, while $M = S \oplus r$ denotes the $r$-level set, forming a double covering of $S$. Due to the large iso-value $r$, the initial $r$-level set starts at a considerable distance from $S$. During the iterative shrinking of the region between the two layers, the opposing effects of distance loss and Laplacian loss frequently result in a significant gap between DCUDF's extracted level set (orange curve) and the actual zero level set, especially in regions of high curvature. DCUDF2 addresses this by incrementally decreasing the effect of the Laplacian loss in areas of high curvature, thereby increasing the influence of the distance loss. The plots on the right display the self-adaptive weights $w_\text{sa}$ for two representative points—one in a low-curvature area and the other in a high-curvature area. This adjustment allows the optimization process to effectively reduce the gap, drawing the orange curve closer to the true zero level set.
• Figure 3: Visual results during the minimization of the accuracy-aware loss function. Initially, uniformly high weights are applied across the entire low-resolution model. As optimization progresses, these weights are dynamically adjusted to focus increasingly on regions requiring finer fitting and refinement. The top row displays the weight values using a heat color map, where warm colors indicate higher weights and cool colors represent lower weights. The middle row showcases the meshes with wireframes, highlighting the introduction of new vertices that provide additional degrees of freedom necessary for better adapting to regions with geometric details. The bottom row illustrates the iterative results of DCUDF, which, using uniform weights, fails to capture the finer geometric details.
• Figure 4: Topology correction can reduce dependency on the optimal setting of the iso-value $r$. In (a), we provide a UDF visualization of a cross-section involving a cylinder and a flat face, which are close to each other. A suitable small $r$, represented by the white iso-curves, accurately separates the cylinder from the plane, yielding an initial mesh $\mathcal{M}$ that preserves the topology of the actual zero level set $\mathcal{S}$. In contrast, an excessively large $r$, shown with orange iso-curves, leads to poor initialization by merging the cylinder and the plane, as illustrated in (b). By clustering the high weight region ($w_{sa}>w_s$), shown with red, we identifies it as cylinder if there are two closed boundary, as shown with yellow line in (b).
• Figure 5: Optimization correction in U-shaped region. Optimizing the $r$ level set towards zero level set may fall into a local optimal solution in a U-shaped region. Although our self-adaptive weights can push the surface a little inward from (b) to (c). As shown in (a), it is difficult to optimize further as the loss gradient is almost perpendicular to the expected optimization direction. By correcting the optimization direction with the mix of gradient and surface normal, we can reach a good performance in (d).