Skip to main content
Back to timeline
arXivSource publication:

Polarization priors replace sparse point-cloud init: Pol2DGS cuts 2D Gaussian surface reconstruction from about 21 minutes to 4.3

Synopsis

Pol2DGS turns multiview polarization observations into dense surface priors, fixes the depth scale with sparse geometric anchors, and places 2D Gaussians anisotropically according to local curvature, so surface coverage exists before optimization starts. Across seven real scenes from PISR and PANDORA, average optimization time is 4.3 minutes versus about 21.3 minutes for PhyGaP, and on three of the four PISR scenes it improves both chamfer distance and F-score while keeping novel-view synthesis quality competitive.

Source-provided article image: RAPID: Reconstruction Accelerated by Polarization-Informed Depth for 2D Gaussian Splatting
arXiv

Interpretation

Model-building time for 2D Gaussian surface reconstruction drops sharply: Pol2DGS averages 4.3 minutes of optimization on the evaluated scenes, against 21.3 minutes for the polarization-based deferred-rendering method PhyGaP. Earlier 2DGS-style methods start from a sparse point cloud and must repeatedly adjust disk positions, orientations and scales plus densify to establish surface coverage; in reflective and weakly textured regions, cross-view color variation gives limited geometric guidance. On seven real captured scenes from PISR and PANDORA, Pol2DGS achieves both lower chamfer distance and higher F-score on three of the four PISR scenes, and its novel-view PSNR and SSIM are best or second-best in most scene comparisons.

Surface coverage is established before optimization begins, with Gaussian positions, orientations and two tangential scales taken directly from the polarization-recovered surface. Polarization gives surface orientation but not position along the viewing ray; under perspective projection normals constrain only log-depth gradients, so integration alone recovers relative depth. Each view's log depth is solved by least-squares integration with sparse geometric anchors, and independent regions lacking anchors receive one reference depth from the visual hull; the per-view depths are then fused with equal weights into a truncated signed distance field to extract the initial surface.

Initial Gaussian density follows surface bending, sparse on flat regions and finer on curved ones, rather than placing one Gaussian per depth sample. A single sampling density creates redundant coverage in low-curvature regions and loses detail in curved ones, and one scale is insufficient when the two principal curvatures differ. A quadratic surface is fitted in local tangent-plane coordinates, the two principal curvatures set sampling spacing along their respective principal directions, and that spacing initializes the two tangential scales; training then adjusts the primitive distribution through insertion, deterministic splitting along the major axis, coverage-constrained merging and opacity-based pruning.

Perspective

Evaluation covers three PANDORA scenes (vase, owl, gnome) and four PISR scenes (BlackLoong, RedLoong, LayingRabbit, StandingRabbit), all real captures, with PISR providing ground-truth meshes from structured-light scanning; inputs require calibrated cameras, polarization observations, foreground and polarization-validity masks, and feature tracks.

Polarization azimuth carries two kinds of directional ambiguity, a π periodicity and the orthogonal directions of diffuse versus specular reflection, which the method suppresses jointly with a periodic angular-difference penalty, an orientation constraint and a smoothness term; ablation shows that removing the extra polarization-normal supervision yields more accurate surface normals on the evaluated scenes, indicating the prior mainly serves initialization rather than full-training supervision.