PISR: Polarimetric Neural Implicit Surface Reconstruction for Textureless and Specular Objects

TL;DR

PISR is presented, a novel method that utilizes a geometrically accurate polarimetric loss to refine shape independently of appearance and leverages the hash-grid-based neural signed distance function to accelerate the reconstruction.

Abstract

Neural implicit surface reconstruction has achieved remarkable progress recently. Despite resorting to complex radiance modeling, state-of-the-art methods still struggle with textureless and specular surfaces. Different from RGB images, polarization images can provide direct constraints on the azimuth angles of the surface normals. In this paper, we present PISR, a novel method that utilizes a geometrically accurate polarimetric loss to refine shape independently of appearance. In addition, PISR smooths surface normals in image space to eliminate severe shape distortions and leverages the hash-grid-based neural signed distance function to accelerate the reconstruction. Experimental results demonstrate that PISR achieves higher accuracy and robustness, with an L1 Chamfer distance of 0.5 mm and an F-score of 99.5% at 1 mm, while converging 4~30 times faster than previous polarimetric surface reconstruction methods.

Figures (7)

• Figure 1: Normal maps of reconstructed surfaces. By leveraging polarization information, our method is able to reconstruct textureless and specular surfaces with fine-grained details.
• Figure 2: Overview of PISR. Pixels for the optimization are sampled in criss-cross patterns at the middle stage of the optimization. The polarimetric loss $\mathcal{L}^\text{p}_\text{pol}$ and the normal loss $\mathcal{L}_\text{normal}$ are used for regularizing the shape independently of appearance.
• Figure 3: Dataset and collection. (a) Objects with ground-truth meshes. (b) Objects without ground-truth meshes. (c) Capture setup.
• Figure 4: Signed error maps. Red and blue mean surface swelling and shrinking respectively. Regions with higher color saturation indicate longer distances between the reconstruction result and the ground truth. Errors are truncated to within $\pm 2$mm for better visualization.
• Figure 5: Normal maps of Dragon B. before and after optimizations. Using neural SDFs to represent shapes allows changes in the shape topology during the optimization.