An arXiv preprint reports that a point-based 3D reconstruction method produced the lowest mean score for mismatch between reconstructed and target surfaces — the metric the study calls symmetric Chamfer distance — in its comparison. It was best overall on three of the five synthetic scenes and best among point-based methods on four; SuGaR was slightly better on LEGO. The ranking is a result from this benchmark, not evidence that the same performance will hold on real captures.
The document is an arXiv preprint, version 1, dated 20 Aug 2026. Across five synthetic objects reconstructed from ten posed views, the reported mean Chamfer distance was 28.5% lower than the strongest point-based baseline. The proposed system used 267 surfels on average, or approximately 161 times fewer primitives.
How the reconstruction works
At the center of the approach is a differentiable point renderer built from compact, opacity-bearing beta surfels. A surfel is a surface element represented as a point; in this formulation, the local opacity-profile shape can also be optimized. The renderer uses an opacity-explicit adjoint light-transport formulation, with gradients for geometry, albedo, opacity and the local profile shape.
The study asks whether evaluating light transport at local surfel coordinates can help a compact reconstruction match posed reference images under known direct illumination while recovering geometry from sparse views. In this setup, illumination is a stated part of the problem rather than a quantity estimated jointly with the scene.
Training combined RGB reconstruction with terms for depth distortion, normal consistency and weak visibility-weighted opacity regularization. The regularization weights were selected empirically and fixed across all reported scenes, and optimization ran for 60K iterations.
The optimized points were converted into meshes by fusing their depth maps at the input camera poses with TSDF, a way of combining depth information into a surface. The evaluation used symmetric Chamfer distance; its directional measures separate accuracy, which captures spurious geometry, from completion, which captures missing regions.
A tightly controlled synthetic test
The benchmark used five isolated objects: Teapot, LEGO, Dragon, Horse and Plant. Each scene supplied ten posed images at 500 × 500 pixels, with known camera calibration and three to four static near-field point lights.
The objects were Lambertian, meaning their surfaces followed a diffuse-reflection model, and appeared against black backgrounds. Reference images used the same direct point-light model as the optimization, with no indirect illumination; Blender Cycles was used as an independent renderer.
Point-based methods shared an initialization of 25 white primitives at the same positions, with an initial scale of 0.1, opacity of 0.5 and vertical orientation. The main point-based comparisons were 2DGS, SuGaR, RadiosityGS and GOF.
This design makes the result a test of sparse-view reconstruction under calibrated, known lighting, rather than a broad test of capture conditions. The evaluation did not include real captures, non-Lambertian reflectance, jointly estimated illumination or large-scale scenes.
The authors state that rendered data, calibration and scene-generation code are released. They also state that the original meshes and textures are not.
What the scores show
Table 1 reports a mean symmetric Chamfer entry of 1.536 for the proposed method, alongside an average of 267 optimized primitives. The preprint reports the number as a point estimate and gives no uncertainty measure for it.
Table 2 reports directional Chamfer accuracy of 0.017 and completion of 0.014. Read together with the metric definitions, those values describe two sides of the reconstruction problem: extra surface on one side and missing surface on the other.
The baseline table also reports different results for two 2DGS schedules: the 7K configuration had mean Chamfer 2.149 with 43.0k primitives, while the 30K configuration had 38.455 with 51.0k. In this evaluation, the shorter configuration had the lower reported entry, so the table does not show a simple link between a longer schedule, a larger primitive count and a lower score.
An ablation on Horse examined the full setup and versions with individual components changed. The full configuration's CD entry was 0.0132; the entries were 0.0162 without depth distortion, 0.0185 without normal consistency, 0.0241 without the weak opacity prior and 0.0278 with fixed beta.
That ablation covers one scene only, so it does not establish whether the same pattern would appear for Teapot, LEGO, Dragon and Plant. More broadly, the preprint reports point estimates without confidence intervals, standard deviations, repeated-run variability or inferential statistical tests.
Mesh quality was assessed after TSDF depth fusion, which means the reported Chamfer values reflect the extraction pipeline as well as the optimized surfel representation. This is another reason to read the scores as results of the full reconstruction pipeline.
The unanswered test
The authors describe the method as a constrained first instantiation of transport optimization with compact beta surfels. The evaluation supports only the narrower claim that this setup produced the reported geometry results in its synthetic, known-lighting benchmark; it does not establish performance beyond that reconstruction task.
Whether the compactness advantage transfers beyond the benchmark remains open. Replication on real captures, with indirect illumination, jointly estimated lighting and non-Lambertian materials, would be needed to assess generalizability; comparisons with matched schedules and broader scene complexity would test whether the reported gap persists.
The Horse ablation leaves a similarly narrow question: whether the reported changes associated with removing regularizers or fixing beta generalize beyond Horse. The current results do not answer it, because the ablation was not reported across all five scenes.
For now, the strongest conclusion is limited but clear: under known direct illumination, on five synthetic isolated objects and ten posed views per scene, the method produced the lowest reported mean surface-mismatch score while using a compact primitive set. Its behavior on real captures, more complex lighting and materials remains untested.
Paper data and sources
Original title: Point-Based 3D Reconstruction from Sparse Views under Known Illumination
Authors: Magnus Kaufmann Gjerde, Joakim Bruslund Haurum, Jeppe Revall Frisvad et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text