iRBSM: A Deep Implicit 3D Breast Shape Model

TL;DR

The paper addresses the need for expressive 3D breast shape models beyond PCA-based approaches, which suffer from occlusion-induced correspondence errors and lack of detail. It introduces iRBSM, an implicit neural representation that models surfaces as the zero-level set of a latent-conditioned signed distance function learned via auto-decoder training on raw 3D breast scans, eliminating the need for rigid/non-rigid registration. The approach yields higher-fidelity surface reconstructions, robustness to sparse, incomplete, and noisy data, and enables a practical single-image to 3D reconstruction pipeline. This work advances clinical and fashion applications by providing a more accurate, flexible, and deployment-friendly breast shape model with publicly available code.

Abstract

We present the first deep implicit 3D shape model of the female breast, building upon and improving the recently proposed Regensburg Breast Shape Model (RBSM). Compared to its PCA-based predecessor, our model employs implicit neural representations; hence, it can be trained on raw 3D breast scans and eliminates the need for computationally demanding non-rigid registration -- a task that is particularly difficult for feature-less breast shapes. The resulting model, dubbed iRBSM, captures detailed surface geometry including fine structures such as nipples and belly buttons, is highly expressive, and outperforms the RBSM on different surface reconstruction tasks. Finally, leveraging the iRBSM, we present a prototype application to 3D reconstruct breast shapes from just a single image. Model and code publicly available at https://rbsm.re-mic.de/implicit.

Figures (4)

• Figure 1.1: Random samples from the proposed iRBSM. Our model is able to generate a large variety of plausible breast shapes, including fine details such as the nipples or belly buttons.
• Figure 1.2: Qualitative results for sparse surface reconstruction. We show surface reconstructions obtained from fitting the RBSM and our model to sparse point clouds. Our model leads to visually more pleasant reconstructions than its PCA-based counterpart.
• Figure 1.3: Qualitative and quantitative results for incomplete (top) and noisy (bottom) inputs. In both scenarios, our model leads to visually more appealing yet plausible reconstructions. This is also reflected quantitatively (right) when compared to learning-free methods, SPSR kazhdan2013 and Matérn weiherer2024. Qualitative results are shown for $r=100$ (top) and $\sigma=10$ (bottom).
• Figure 1.4: Single image to 3D using the iRBSM. We employ our model to 3D reconstruct breast shapes from single images by first estimating depth using a state-of-the-art monocular depth estimation model yang2024 and then fitting the iRBSM to the unprojected depth map. This yields high-quality surface reconstructions with accurate normals in seconds and without special hardware.