Wires, fence strands, bicycle spokes, plant stems and truss members are among the first things to disappear when a 3D scene is rebuilt from only three to eight photographs. The cause is simple. Current methods fit the scene by matching rendered images to the photos pixel by pixel. A wire covers a few dozen pixels, while a wall covers hundreds of thousands, so the fit can ignore the wire at almost no cost. With few views, nothing else rescues it. Tie points from the camera-calibration step rarely land on wires. Monocular depth estimators smooth over them. Densification heuristics either never create geometry there or join neighbouring strands into a sheet.
We propose to treat thin structures as their own kind of object: 3D curves with a radius. Each curve is fitted to 2D curve detections across the sparse views, using a geometric consistency loss rather than a photometric one. Bulk surfaces stay as ordinary 3D Gaussian splats, the explicit point-like blobs used in 3D Gaussian Splatting (3DGS). The two-sentence version is this. A thin structure that is nearly invisible to a pixel-averaged loss still shows up clearly as a line in each image, and a 3D curve is pinned down by two such lines whenever the views are not degenerate for that curve. We build a hybrid splat-plus-curve reconstruction around this fact and use the degeneracy condition to decide which few views a thin-structure scene actually needs.
Core Claim
Under sparse angular sampling (three to eight posed RGB views), the missing ingredient for thin structures is a primitive whose multi-view consistency is enforced on its 2D projected shape, not a stronger generative prior. Concretely, the curve primitive is a polyline of 3D control points with a per-point radius. It is rendered as a chain of elongated Gaussians whose pixel coverage is computed from the true radius. A 0.3-pixel wire is therefore drawn as opaque geometry covering 30% of a pixel, not as a semi-transparent cloud. The primitive is supervised by two losses.
The first is a curve reprojection loss: a two-sided distance between the projected centreline and 2D ridge detections in every view.
The second is a coverage loss modelled on the symmetric-difference loss of CrossSDF. That paper fits signed distance fields to sparse 2D medical cross-sections and penalises only inside/outside disagreement. It also concentrates samples inside thin contours so that tiny structures carry equal gradient weight. We move both ideas from axis-aligned slices to projective camera views.
The prediction is that this preserves thin-structure topology at three views, where splat-only methods lose it. It should do so without hurting whole-image quality.
Why this is not already done
The closest works each cover two of the three directions. CurveFusion fuses 1D curve skeletons from dense handheld RGBD video and reaches 2 mm wires. It needs continuous depth streams, fails on objects that mix wires and surfaces, and fails near walls. We keep its curve-skeleton representation but drive it from sparse RGB and embed it in a hybrid scene.
Geo-Neus recovers thin handles and stems by anchoring an implicit surface on calibration tie points and enforcing patch photometric consistency. It needs enough views for those tie points to exist on the thin parts, and that assumption fails at three views.
DeepView renders railings and fences well from a 16-camera rig using layered planes. It is limited to forward-facing setups and produces no metric geometry.
Among sparse-view splatting methods, FSGS adds Gaussians at midpoints between neighbours. The research file notes this tends to merge parallel wires into sheets. IPSM shows that score distillation fails under sparse views and rectifies warped pseudo-views with an inpainting model. GSFixer repairs sparse 3DGS renders with a video diffusion model conditioned on geometry tokens.
These three papers motivate the gap rather than close it. The two generative methods can only sharpen structure the initial geometry already contains, and neither represents a filament as a connected object. Our proposal adds no generative prior. The new elements are a curve primitive, a curve-specific consistency loss, and a view-conditioning criterion.
Research Program
The pipeline has three stages, and the third connects back to view selection.
Stage 1: seed curves. Thin-structure maps are extracted per view with an off-the-shelf thin-object segmenter, such as RT-DLO for linear objects or the trunk-versus-structure split in UDUN. These are combined with a classical ridge filter so that the method does not depend on one learned detector. 2D curves are matched across views along epipolar lines and triangulated into seed polylines. This replaces both calibration tie points and midpoint unpooling as the source of thin geometry.
Stage 2: joint optimisation. Splats handle bulk surfaces and curves handle filaments. Each pixel's colour composites both. The curve terms are added to the usual photometric and depth-correlation terms. Pixels claimed by a curve are excluded from splat densification, so the splats cannot bridge the gap between two curves.
Stage 3: degeneracy analysis. A curve point $X$ with unit tangent $t$, seen from cameras centred at $c_1$ and $c_2$, has its position along the curve undetermined whenever $t$ lies in the epipolar plane. We use the conditioning score
$$\kappa(X,t;c_1,c_2) = \left| t \cdot \frac{(c_2-c_1)\times(X-c_1)}{\|(c_2-c_1)\times(X-c_1)\|} \right|,$$
which is 0 for a degenerate pair and 1 for the best case. A curve is well determined by a view set if some pair in the set gives high $\kappa$ along most of its length. This turns the "multi-view multiplicity deficit" from the UAV next-best-view study into a concrete requirement. That study found that planning views for surface coverage alone dropped coverage to 88% on a steel truss bridge, compared with 97.5–99% for dense flight grids. For thin structures, the requirement is not coverage but transverse baselines.
We will use $\kappa$ in two ways. First, as a diagnostic: does reconstruction error on held-out curves track $\kappa$? Second, as a greedy selection rule: pick $k$ views from a candidate pool to maximise the worst-case $\kappa$ over the seed curves of a coarse first pass.
Baselines
The naive reference is vanilla 3DGS on the same sparse views. IPSM reports 14.97 dB PSNR for it on LLFF with three views, and GSFixer reports 13.72 dB on DL3DV with three views. These numbers frame how far the field already is from it.
FSGS is the most direct competitor on geometry construction, because its unpooling is exactly the mechanism we replace. It reports 0.652 SSIM on LLFF with three views.
IPSM is the strongest prior-based method with reported numbers on our main benchmark. It reports 20.44 dB PSNR on LLFF with three views, essentially tied with FSGS at 20.43 dB. It also claims recovery of thin leaves and petals, so it is the bar we must beat on thin regions.
GSFixer is the strongest generative repair method. It reports 16.21 dB on DL3DV with three views and a 2.16 dB gain over GenFusion on DL3DV-Res. We will run it on the DL3DV scenes containing railings and foliage.
All baselines get the same views, the same camera poses, and the same monocular depth network. Each gets the same per-scene iteration budget, measured in GPU-minutes. Each also gets a hyperparameter sweep of equal size on a held-out validation scene.
The key ablation is "mask-weighted splats." It uses identical 2D thin-structure maps and the identical coverage and reprojection losses, applied to ordinary Gaussians with standard densification instead of curve primitives. A second ablation removes epipolar seeding and initialises curves randomly.
Whole-image PSNR on these benchmarks is dominated by bulk regions. Beating a baseline therefore means two things. On whole-image metrics, staying within about 0.2 dB of IPSM. On thin-structure-masked metrics and topology, showing a clear gain. We found no reported result on Lacuna for any method on masked thin-region metrics, so all such numbers will be produced by us under one protocol.
Why not just…?
Why not just reweight the photometric loss on thin-structure pixels in FSGS or IPSM? This is the most serious threat. It addresses the pixel-imbalance argument directly and costs one line of code. What it leaves alone is how geometry gets created and how it is shaped. Reweighting cannot spawn Gaussians where none exist, and midpoint unpooling will still fill the gaps between strands. The mask-weighted-splats ablation tests exactly this. If it matches the full method on curve completeness and on false bridges between strands, the curve primitive is unnecessary and we will report that.
Why not just initialise from a feed-forward geometry network such as VGGT? GSFixer already conditions on VGGT tokens. The research file notes that when the initial sparse geometry omits a sub-pixel wire, those tokens cannot anchor it. We will measure directly how often VGGT-style point maps contain points on thin structures in our scenes. If they reliably do, they become a better seeding stage for our method, not a replacement for it.
Why not just use a stronger diffusion prior to hallucinate missing filaments? IPSM shows that score distillation under sparse views drifts toward blur and floaters. Neither IPSM nor GSFixer represents connectivity, so a hallucinated wire can be broken or inconsistent across views. Our topology metrics on IPSM and GSFixer outputs will show whether this is a practical problem or only a theoretical one.
Why not just capture more views? Sometimes that is the right answer, and Stage 3 says when. The $\kappa$ analysis predicts which scenes need an extra view and from which direction. We test this by comparing $\kappa$-selected view sets against uniform and coverage-selected sets of the same size.
Proposed Experiments
Real scenes with thin structures. These are LLFF with three views (fern, flower, orchids), the Mip-NeRF 360 bicycle scene, and Spaces scenes with railings and chain fences. Mip-NeRF 360 itself notes that its spokes remain blurred even with dense capture. For the bicycle and Spaces scenes we subsample to three, five and eight views.
Geometry. We use DTU with three views, with Chamfer distance restricted to ground-truth points within thin regions.
Synthetic Blender scenes with exact curve ground truth. These include wire sculptures, a chair with spindles, a fence in front of a wall, and a truss lattice modelled after Bridge B. The fence-in-front-of-a-wall scene specifically probes the near-surface case where CurveFusion fails.
Metrics. Beyond PSNR, SSIM and LPIPS on full and thin-masked regions, we measure four things. Curve completeness and precision come from skeleton Chamfer distance at fixed thresholds. Junction accuracy follows CurveFusion's protocol. A bridging rate counts the fraction of ground-truth gaps between strands that are filled. Connected-component error is computed per filament.
Decisive checks. There are four.
The first is whether full-method thin-region completeness clearly exceeds both IPSM and the mask-weighted-splats ablation at three views, while whole-image PSNR stays within a small margin of IPSM.
The second is whether the FSGS bridging rate is substantially reduced. This also tests the research file's claim about FSGS bridging, which so far is stated rather than measured.
The third is whether per-curve error correlates negatively with best-pair $\kappa$ on synthetic scenes. If it does not, the degeneracy story is wrong and Stage 3 is dropped.
The fourth is whether $\kappa$-selected five-view sets beat uniform five-view sets on curve completeness on the truss scene.
A negative result on the first check against the ablation would redirect the project toward loss design rather than representation.
References
- CrossSDF: 3D Reconstruction of Thin Structures From Cross-Sections
- CurveFusion: Reconstructing Thin Structures from RGBD Sequences
- Geo-Neus: Geometry-Consistent Neural Implicit Surfaces Learning for Multi-view Reconstruction
- DeepView: View Synthesis With Learned Gradient Descent
- FSGS: Real-Time Few-Shot View Synthesis Using Gaussian Splatting
- How to Use Diffusion Priors under Sparse Views?
- GSFixer: Improving 3D Gaussian Splatting with Reference-Guided Video Diffusion Priors
- Autonomous UAV 3D Reconstruction using Prediction-Based Next Best View
- Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields