Seeing Only What Was Photographed: Support-Aware Sparse-View Splatting for Scoring Taxonomic Characters

Taxonomists sort organisms into species and work out their family trees by scoring physical traits: the number of segments in an antenna, whether a leg branches, the shape of a joint. These traits are called characters. More and more of this scoring happens from digital 3D models of museum specimens instead of the fragile specimens themselves. Fast 3D reconstruction from a handful of photographs is now practical. One example is 3D Gaussian splatting, which represents an object as many small colored, semi-transparent blobs that a graphics card can draw hundreds of times per second. The danger is that a model built from few photographs fills the unseen parts of the object with invented geometry. It may add floating specks, flatten a leg into a card, or produce a smooth, plausible surface where the real specimen had a spine. A taxonomist rotating that model in real time cannot tell which parts were photographed and which were guessed. Any character scored from a guessed region is a fake data point. In a field where the central difficulty is already telling true shared ancestry from look-alike traits (homoplasy), fake data points can move branches of the tree.

This note proposes to make the reconstruction report, for every pixel it draws, how many input photographs actually support that surface, at no meaningful cost to drawing speed. Characters that fall in unsupported regions are then entered as "unknown" rather than scored from hallucinated geometry. The research question is not whether sparse-view reconstruction looks good. It is whether character matrices scored from sparse-view reconstructions, with or without this support signal, yield the same evolutionary trees as matrices scored from the physical specimens.

Core Claim

Image-quality metrics such as peak signal-to-noise ratio (PSNR) are the wrong target for specimen digitization. A reconstruction can score well on held-out views while inventing exactly the fine structures that carry taxonomic signal. We claim three things. First, sparse-view artifacts produce systematic character-scoring errors that are measurable on a benchmark built against physical-specimen scoring. Second, a per-blob support count can be blended into the image exactly like color and drawn in the same rasterization pass. This count separates trustworthy from untrustworthy regions well enough that masking unsupported characters restores tree topology. Third, it does this nearly as cheaply as the base renderer, whereas ensemble-based uncertainty multiplies cost. Latency is therefore treated as a constraint: the support signal is only useful if it can be shown live during interactive scoring.

What Exists and What Is Missing

Two separate literatures meet here without touching. On the vision side, several methods attack sparse-view artifacts while keeping real-time rendering, all evaluated on generic scenes. FSGS inserts new blobs between existing neighbors, adds synthetic in-between views, and supervises depth with a pretrained monocular depth network. DNGaussian anchors blob positions to normalized monocular depth while freezing blob shapes to avoid blur. HeroGS trains three blob sets in parallel and prunes blobs that lack neighbors in the others. Few-Shot Novel View Synthesis Using Depth Aware 3D Gaussian Splatting adds depth loss and disables opacity-based pruning. Earlier, RegNeRF regularized depth smoothness and patch appearance from unobserved viewpoints. All of these make artifacts less visible, and that is precisely the concern. Smoothness priors and monocular depth fill unobserved regions with confident-looking but invented surfaces, and none of these methods tells the user which surfaces those are. The file also notes that monocular depth networks are biased on macro, close-up imagery. That bias is the regime of insect specimens.

Two feed-forward methods supply the mechanism we reuse. Stereo Radiance Fields predicts density from cross-view feature agreement, so points without stereo correspondence receive low confidence. MVSplat ties each blob's opacity to multi-view matching confidence. Both use agreement across views to suppress geometry. Neither exposes that agreement as a separate, renderable quantity, and both fold it into opacity, where it is entangled with true transparency. We keep it as its own channel.

On the taxonomy side, digitization work motivates the problem without solving it. Towards High-Throughput 3D Insect Capture built a robotic pan-tilt rig with encoder-derived camera poses. It reports that focus stacking at each angle makes many-view capture slow, which is exactly why few views are attractive. It used classical photogrammetry and never asked whether reconstructions preserve characters. Primitive-based 3D structure inference from a single 2D image for insect modeling fit cylinders and ellipsoids to a single photograph. This is the extreme version of guessing unseen geometry from a prior, and it yields "visually plausible" models that cannot capture wings or mouthparts. Urokodia aequalis shows that 3D characters recovered by micro-CT, fed into a 71-taxon, 95-character matrix, can relocate a fossil to the base of Artiopoda. That result tells us both that 3D characters move trees and that micro-CT lacks color. Prabhat et al. demonstrate the stakes of homoplasy in a parsimony setting from caliper measurements alone. To our knowledge no work in the file measures character-level fidelity of sparse-view reconstructions, and none attaches renderable support to real-time splats.

Research Program

The method adds one quantity to a sparse-view Gaussian splatting model built on DNGaussian or FSGS. For each blob $i$, the support $s_i$ is the number of input views $v$ in which the blob is both visible and consistent:

$$s_i = \sum_{v=1}^{V} \mathbb{1}\left[T_{i,v} > \tau\right]\,\mathbb{1}\left[\,\lVert f_v(\pi_v(\mu_i)) - \bar f_i \rVert < \epsilon\,\right].$$

Here $V$ is the number of input photographs, $T_{i,v}$ is the light transmittance reaching blob $i$ from camera $v$ (so a blob hidden behind others does not count), $\tau$ is a visibility threshold, and $\mu_i$ is the blob center. The projection $\pi_v$ maps a 3D point into image $v$, $f_v$ is a pretrained image-feature map, $\bar f_i$ is the mean feature across the views that see the blob, and $\epsilon$ is a consistency tolerance.

A surface needs at least two consistent views to be triangulated rather than guessed. Rendered support below 2 therefore marks geometry that came from priors such as monocular depth, smoothness, or densification, not from observation. The support is alpha-blended into a single extra image channel in the same pass as color. The added cost is one scalar per blob, so frame rate should stay close to the base method. We will measure this rather than assume it.

The program has three parts. The first is a benchmark. We will capture insect specimens on a calibrated pan-tilt rig of the kind described by Nguyen et al., with dense focus-stacked views. We will then subsample to 2, 3, 6 and 12 input views and keep held-out views for evaluation. We will choose taxa from Pentatomidae and Arctiinae, where mitogenomic trees exist (Pentatomidae, Arctiinae). Experts will score a morphological character list three ways: under the microscope from the physical specimen, which is the reference; from the sparse input photographs alone; and from each reconstruction. As a geometry-only controlled testbed, we will texture and render surfaces from the published Urokodia micro-CT scans. This gives a fossil case with exact 3D ground truth.

The second part is the support-aware reconstruction and a scoring interface. The interface overlays support while the expert rotates the model, and its protocol converts characters whose diagnostic region has support below 2 into missing data. The third part is downstream analysis. We will build parsimony and Bayesian trees from each matrix and compare their topology with the physical-scoring tree.

Baselines

The naive reference point is an expert scoring characters directly from the same 2, 3 or 6 sparse photographs, with no 3D model. This is what taxonomists do today. If reconstruction does not beat it on character accuracy, 3D adds nothing at these view counts. Both conditions use the same specimens, the same photographs and the same scorers, with scoring order randomized across conditions to limit memory effects.

The strongest existing per-scene method is DNGaussian, with FSGS as a close second. DNGaussian reports 300 FPS rendering, 3.5-minute training and 2 GB of memory on standard benchmarks. FSGS reports 203 to 458 FPS. HeroGS reports PSNR 18.78 on LLFF with 2 views, outperforming FSGS. These are the competitors because they are the methods a digitization lab would adopt for fast, few-view, interactive models. We will run them with their published settings on identical inputs, with equal training iterations and the same hyperparameter search budget applied to our variant. For the strongest feed-forward competitor, MVSplat reports PSNR 26.39 on RealEstate10K. Since it cannot be retrained on our small specimen set, we will run it zero-shot and treat it as a reference. We found no reported result on Lacuna for any of these methods on insect or other biological specimen imagery, so all specimen numbers will be new. Beating them means higher character-scoring accuracy on scored characters at equal or better coverage, and smaller tree distance to the physical-scoring tree, at a frame rate within roughly 10% of the base method.

Ablations isolate the key ingredient. A visibility-only support count, without the feature-consistency term, tests whether consistency matters. A random-masking control removes the same fraction of characters as support masking, so we can check that any tree improvement comes from where we mask and not simply from having fewer characters. Running each base method with and without monocular depth supervision tests whether depth priors are the main source of confident hallucination on macro imagery.

Why not just…?

Why not just photograph more views? If dense capture were cheap, support masking would rarely trigger. Nguyen et al. report, however, that each angle at macro scale requires focus stacking, and that their light-field shortcut costs the resolution needed for fine structures such as setae and sensilla. Our view-count sweep answers the question directly by finding the number of views at which character accuracy saturates. If 12 views already match physical scoring at acceptable capture time, we will report that as the practical recommendation. The masking contribution then matters only below that threshold.

Why not just use micro-CT? It gives true 3D, as Urokodia shows. It is slow and lab-bound, though, and records density, not the color and pigmentation that many insect characters depend on. We use it as ground truth for geometry, not as the competitor for routine capture.

Why not just estimate uncertainty with an ensemble or several training seeds? This is a real threat. Disagreement among independently trained models may flag hallucinated regions as well as support counts do. It costs a multiple of training and rendering, however; HeroGS's three parallel fields triple the cost. We will run a 3-member ensemble as an uncertainty baseline and compare the area under the curve for detecting misscored characters. If the ensemble matches support counts, our contribution reduces to doing the same thing at single-model cost, and we will report it that way.

Why not just tune the regularizers until artifacts disappear? Fewer visible artifacts is not the same as fewer wrong characters. Smoothness and depth priors may convert a floater into a plausible but false surface. The comparison of character accuracy against image-quality scores across regularizer settings tests whether these two measures diverge.

Proposed Experiments and Decisive Checks

We will evaluate held-out-view PSNR, SSIM and LPIPS (a perceptual similarity metric), frame rate with and without the support overlay, character accuracy against physical scoring, coverage (the fraction of characters scored rather than masked), and Robinson–Foulds distance between trees. Robinson–Foulds distance counts the branch groupings that differ between two trees.

The first decisive check is whether, among characters scored from reconstructions, error is concentrated in regions with rendered support below 2. A strong separation, measured as detection AUC well above that of ensemble variance or of a plain visibility count, would validate the signal. If misscored characters are spread evenly across support levels, the idea fails.

The second check is whether masking unsupported characters reduces tree distance to the physical-scoring tree more than random masking at equal coverage. We also ask whether it reduces conflict with the independent mitogenomic topology relative to unmasked reconstruction scoring. The physical-scoring matrix sets the floor for morphology-versus-molecule conflict that homoplasy alone produces.

The third check is the cross-plot of PSNR against character accuracy across all methods and view counts. A weak or inverted relationship would establish that image metrics are the wrong target for this application. EvalMVX (EvalMVX) offers a template for reporting results across capture setups.

Finally, on the textured Urokodia surfaces, we will re-score the micro-CT-derived characters from sparse-view reconstructions. We then ask whether the fossil's basal placement in the 95-character matrix survives with and without support masking. This gives a direct test of whether reconstruction artifacts can manufacture or erase a published phylogenetic conclusion.