Cameras that see through fog, murky water or smoke almost always assume that light bounces off the medium at most once on its way to the sensor. This "single scattering" assumption keeps the math simple. It underlies most underwater 3D reconstruction and dehazing methods, including recent neural scene models deployed on real ocean dives. In thick media, though, light bounces many times. Images blur, a diffuse glow builds up, and reconstruction methods start to explain this glow as fake geometry or shading. Physically complete alternatives exist: differentiable renderers that solve the full equation for light transport through a medium. They have been validated only on synthetic clouds, or on real aerial imagery that has no ground truth. As a result, nobody knows at what turbidity the cheap models actually fail, or whether the expensive ones actually fix the problem on real data.
We propose to answer that question experimentally. We will build a water tank with milk titrated in to set the scattering strength, place objects of known shape in it, and measure the medium's optical properties independently of any reconstruction. We then run a ladder of image-formation models on the same images across a sweep of turbidity, from strict single scattering, to single scattering plus a 2D blur kernel, to a full Monte Carlo multiple-scattering renderer. The output is an empirical regime map that says, as a function of optical thickness, which model is needed to recover geometry to a given accuracy. The second contribution is a validation protocol. A model counts as correct only if it predicts images under lights it was never fit to, using medium parameters it was never allowed to tune.
Core Claim
Our claim is that model choice is a measurable function of the medium, and that a predictive test with held-out lights is the right referee. Optical thickness $\tau = \sigma_t L$ is the extinction coefficient times the camera-to-object path length, so it counts expected collisions per path. As $\tau$ grows, there is a threshold beyond which a single scattering model plus a 2D blur kernel stops matching held-out images and starts producing biased depth and normals. Beyond that threshold, a full multiple-scattering renderer with independently calibrated medium parameters restores both. We expect the threshold to depend on how forward-peaked the medium's scattering is, not on $\tau$ alone. This is exactly the tension the literature leaves open. Multiple-Scattering Microphysics Tomography argues that full 3D radiative transfer is necessary in dense clouds, where it cut synthetic mass error from −55% to −8%. Photometric Stereo in a Scattering Medium and Unsupervised water scene dehazing network using multiple scattering model treat multiple scattering as a single convolution kernel and report good results. Both positions can be right in different regimes, and no experiment has located the boundary.
Relation to Prior Work
The closest physical testbeds are the milk-water tanks of Structured Light in Scattering Media and Photometric Stereo in a Scattering Medium. The first derived a single-scattering model of backscatter and attenuation and showed that five light sources are needed for photometric stereo in a medium. The second added a spatially invariant deblurring kernel and fluorescence filtering. These papers supply our experimental method, and we reuse the titration idea and the box-in-a-tank refraction calibration. Both papers explicitly stop where multiple scattering dominates. Neither measured medium parameters independently, and neither tested a full transport solver.
On the modeling side, Light Transport-aware Diffusion Posterior Sampling for Single-View Reconstruction of 3D Volumes embeds a differentiable multiple-scattering renderer in a diffusion sampler but is tested only on simulated clouds. Multiple-Scattering Microphysics Tomography inverts full 3D transport on real AirMSPI aerial images but has no in-situ ground truth. These papers motivate the proposal, because they show the full model matters in simulation. They also supply the gradient machinery we reuse.
SeaThru-NeRF and Decoupling Scattering are neural radiance fields, meaning neural networks that represent a scene as a 3D field queried along camera rays. Both were deployed on real dives and both assume single scattering. They are the practical baselines the regime map is meant to inform. Differentiable Direct Volume Rendering constrains the design, because its memory-efficient gradient trick does not extend to scattering, so we cannot simply borrow it. SEAGLE is the one multiple-scattering inversion validated on a real benchmark with known targets, the Fresnel microwave data. It handles discrete objects in air rather than a continuous turbid volume. Our testbed is the optical, continuous-medium counterpart of that benchmark.
Research Program
Testbed. The testbed is a glass tank with a refraction-calibration box. It holds 3D-printed matte objects with known CAD geometry, registered to the camera by fiducials scanned in clear water. Six or more calibrated point lights are mounted at known positions, with a fixed multi-view camera rig. Milk is titrated in a geometric series of concentrations, and grape juice is added in a second series to vary absorption independently, following Photometric Stereo in a Scattering Medium.
Medium calibration. At each concentration we measure the medium with instruments that are separate from the reconstruction. Collimated laser transmission over two path lengths gives the extinction coefficient from ballistic attenuation. A small goniometric measurement of a thin sample gives the phase function, summarized by its asymmetry parameter $g$ and fitted with a Henyey–Greenstein or two-lobe form. Total diffuse reflectance from a deep cell constrains the albedo. These measured values define $\tau$ for each condition and are frozen for the predictive tests.
Model ladder. All models share the same surface representation, a signed-distance field with an albedo map. They differ only in the image-formation model. The first rung is the single-scattering model used by SeaThru-NeRF. The second adds a depth-invariant 2D blur kernel, as in Murez et al. and the unsupervised dehazing network. The third makes that blur kernel depth-dependent. The fourth is a differentiable volumetric path tracer that tracks multiply scattered paths through the measured medium, as in Pattanaik et al.'s Monte Carlo participating-media method, with gradients computed as in the microphysics tomography work. Each rung is run in two modes. In one, medium parameters are fixed to the calibrated values. In the other, they are estimated jointly with geometry.
Validation protocol. Every model is fit using a subset of lights and views. It is then asked to predict raw images under held-out lights and viewpoints. This test targets the failure mode identified in the PDPS and DecoNeRF papers, where shading gets painted onto geometry. A model that bakes scattering into shape can fit its training images but will mispredict under a new light.
Baselines
The naive reference is classical photometric stereo with no medium model. Photometric Stereo in a Scattering Medium reports 14% depth error and 26° normal error for this method in highly turbid water. We rerun it on our tank so that its curve anchors the bottom of the map.
The strongest existing physical method on this exact question is the same paper's deconvolution-plus-fluorescence pipeline, which reports 4% depth error and 13° normal error in its tank. We rerun it with fluorescent-coated copies of our objects, and separately with deconvolution only on uncoated objects, because the fluorescence advantage is unavailable to every other method. The single-scattering structured-light model of Structured Light in Scattering Media reports RMS error of about 2.1% at moderate and about 6.3% at high scattering. We reimplement it as the strict single-scattering rung.
For neural methods, we run SeaThru-NeRF and DecoNeRF on our multi-view captures with known poses, which removes the COLMAP failure mode. SeaThru-NeRF reports about 0.8 dB PSNR over Mip-NeRF 360 on real dives, and DecoNeRF reports 21.42 versus 15.88 PSNR on hazy scenes. Neither paper reports geometry error against measured ground truth, so there is no reported result found on Lacuna for neural methods' depth error in a calibrated turbid tank.
For full multiple-scattering inversion on real optical data with ground truth, no reported result found on Lacuna either. PDPS's 38.46 PSNR is on synthetic clouds only.
All methods see identical images, lights and poses. Each gets the same hyperparameter search budget, tuned on the lowest-turbidity condition only, and the same wall-clock cap per reconstruction. The path tracer is also given a compute-matched variant with fewer samples, so that its cost can be reported honestly. Beating the strongest baseline means two things at turbidities where it degrades. First, lower depth and normal error than deconvolution-only with no fluorescence. Second, matching or beating deconvolution with fluorescence. Both must come with lower held-out-light image error. The key ablation is the full path tracer with jointly estimated rather than calibrated medium parameters. If that ablation performs as well, independent calibration is unnecessary and the protocol simplifies.
Why not just…?
Why not just keep the single-scattering model and let a flexible neural network absorb the mismatch? Networks such as the medium branch of SeaThru-NeRF can fit training images well, but fitting is not evidence of correct geometry. Decoupling Scattering documents exactly this kind of trivial solution, in which haze gets baked into solid geometry. Our held-out-light test is designed to catch it. If the neural single-scattering model matches the path tracer on geometry and on held-out-light prediction at high $\tau$, we will report that the extra physics is unnecessary.
Why not just add a 2D blur kernel, as Murez et al. and the dehazing network do? This is the real threat. It is cheap, and it may suffice up to a large $\tau$, particularly when forward scattering is strong and milk-like. We do not argue it away. The regime map is designed to locate where the kernel fails, if it fails anywhere inside the tank's range. A null result, in which the kernel matches the path tracer everywhere we can measure, would itself be a useful finding for practitioners. The depth-dependent kernel rung separates "blur that varies with depth" from "true multi-bounce transport."
Why not just run this study entirely in simulation with a path tracer? Simulation assumes the very things in question: the phase function, the tank-wall interreflections, and the sensor response. The SeaThru-NeRF dives and AirMSPI show how real conditions differ from synthetic phantoms, and RESIDE shows that synthetic-benchmark ranking need not transfer to real data (RESIDE). We will run the matched simulation anyway. The gap between simulated and measured regime maps is a reported result in its own right.
Why not just gate scattered light out in hardware? Fluorescence only works for fluorescent targets, and it removes backscatter but not forward-scatter blur. Photometric Stereo in a Scattering Medium still needed deconvolution on top of it. We include fluorescence as a baseline rather than as the solution.
Decisive Checks
The first check is whether the per-model curves of depth and normal error against measured $\tau$ cross at all within the tank's range, and where. The second is whether held-out-light image error predicts geometry error. If it does, the protocol can be used where geometric ground truth is unavailable, such as in the field. The third is whether the jointly estimated medium parameters agree with the independent laser and goniometer measurements. Disagreement that grows with $\tau$ would show that self-calibrating methods are unidentifiable there, which is the parameter ambiguity flagged across the validation literature. The fourth is whether repeating the sweep with a weakly forward-scattering additive moves the crossing point. That would confirm that the threshold depends on $g$ and not only on $\tau$. The capture set, calibration data and CAD will be released as an optical, continuous-medium analogue of the Fresnel microwave benchmark.
References
- Multiple-Scattering Microphysics Tomography
- Photometric Stereo in a Scattering Medium
- Unsupervised water scene dehazing network using multiple scattering model
- Structured Light in Scattering Media
- Light Transport-aware Diffusion Posterior Sampling for Single-View Reconstruction of 3D Volumes
- SeaThru-NeRF: Neural Radiance Fields in Scattering Media
- Decoupling Scattering: Pseudo-Label Guided NeRF for Scenes with Scattering Media
- Differentiable Direct Volume Rendering
- SEAGLE: Sparsity-Driven Image Reconstruction Under Multiple Scattering
- Computation of global illumination in a participating medium by monte carlo simulation
- RESIDE: A Benchmark for Single Image Dehazing