A single-photon lidar fires laser pulses and records the arrival time of individual returning photons. At very low light, a reconstruction depends on two things: where the sensor stands and how long it stays. Each photon is a random count, so depth precision on a surface patch improves only as the square root of the photons that patch returns. How many photons a patch returns is set by geometry. Range, viewing angle relative to the surface, reflectivity and occlusion together decide it. A view that grazes a surface, or sits far from it, wastes most of its exposure.
Current methods treat these two choices separately. Geometry-aware planners choose viewpoints from curvature or surface normals and assume every captured image is clean. Photon-starved reconstruction methods model the counting noise carefully but use fixed, hand-chosen viewpoints. This note proposes one planner that chooses both the next sensor pose and the exposure time at each pose. Its objective is the information a Poisson count actually carries about surface position. That information is computed in closed form from the current shape estimate, so the planner never takes numerical derivatives of noisy geometry.
Core Claim
Under a fixed total photon (or time) budget, the useful planning signal for low-light lidar reconstruction is the photon-weighted information about each surface patch's position. It is not curvature, and it is not generic model uncertainty. For a patch $s$ seen from view $v$ with exposure time $t_v$, the expected signal count scales as $t_v \rho_s \cos\alpha_{vs} / r_{vs}^2$. Here $\rho_s$ is reflectivity, $\alpha_{vs}$ is the angle between the viewing ray and the surface normal, and $r_{vs}$ is the range. Timing error along the ray projects onto the surface normal with a factor $\cos\alpha_{vs}$. The Poisson information about the patch's position along its normal is therefore approximately
$$ I_{vs} = t_v \, \frac{\rho_s \cos^3\alpha_{vs}}{r_{vs}^2 \, \sigma_p^2} \, \gamma_{vs}\,\mathbb{1}[\text{visible}], $$
where $\sigma_p$ is the laser pulse's temporal width converted to distance and $\gamma_{vs} \le 1$ discounts for background photons. This form follows from the standard result that a Poisson time-of-arrival estimate has variance $\sigma_p^2/N$ for $N$ signal photons.
Information is linear in exposure time, so splitting a fixed budget $T$ across already-chosen views is a concave problem:
$$ \max_{t \ge 0,\ \sum_v t_v = T} \ \sum_s \log\Big(\textstyle\sum_v I_{vs}\Big). $$
This is the classical "maximize the log-determinant of the information" design, here applied per patch. We claim three things. First, a sequential planner can do two steps in a loop: choose the next pose by the largest gain in this objective, then re-solve the exposure split. Second, this loop beats uniform orbits by a margin that grows as photons get scarcer. Third, it beats curvature-driven planning, which at low counts chases false edges created by noise.
Relation to Prior Work
The reconstruction engine is taken from Transientangelo. It fits a neural signed distance function (a network whose zero level set is the surface) directly to photon-arrival histograms from 2 to 5 lidar views, down to 10 photons per pixel. Its reflectivity loss already estimates $\rho_s$, which our photon-yield model needs. Its viewpoints are fixed and its exposure is uniform. Our contribution is the planner that chooses both.
Path Matters motivates the work. It shows that sparse-view Gaussian splatting collapses when high-curvature regions are under-sampled, and it plans smooth B-spline camera paths toward curvature. It also constrains the work. Its curvature signal is a second derivative of estimated depth, and second derivatives are exactly what shot noise corrupts. Photon-Starved Scene Inference documents this failure: standard features break down below one photon per pixel. Our objective uses only first-order geometry (the normal and the range) and the count model, so it avoids differentiating noise twice.
Uncertainty-driven view planning supplies the sequential greedy information-gain template. It cut the views needed for sonar hull inspection from 38 to 8. Its observation model, however, is based on SURF features (a classical keypoint detector) rather than photon counts, and it never allocates exposure.
Beam orientation optimization is the closest precedent for choosing angles and per-angle intensity together. It uses group-sparse optimization, and 6 optimized radiotherapy beams outperformed 11 equiangular ones. We borrow the joint angle-and-intensity framing, but our forward model is Poisson sensing rather than dose deposition.
Normal-facing heuristics also inform the pose search. Embodied Crowd Counting flies a drone along surface normals, and Learning Virtual View Selection selects views in a combined position-and-normal space. Our $\cos^3\alpha$ term recovers "face the surface" as a consequence of the count model rather than as an assumption. It also says how much a grazing view is worth, which those heuristics do not.
Research Program
We will build a simulator that renders time-resolved photon histograms from arbitrary poses on the scenes underlying the Transientangelo simulated benchmark. Candidate poses lie on a sphere of camera positions around each object. We will reuse Transientangelo's forward model so that rendered data match its training distribution.
The planning loop runs as follows. Start from one or two seed views at low exposure. Fit the signed distance function. Extract surface patches with normals, ranges and reflectivity. Score candidate poses by the gain in the log-information objective, using visibility traced through the current surface estimate. Add the best pose. Re-solve the exposure split by projected gradient ascent. Capture by thinning the photon stream to the allocated time, then repeat until the budget is spent. An optional travel-length penalty in the pose score gives the smooth paths that Path Matters found necessary.
We will also check calibration. The variance predicted by the inverse information should match the empirical error of patch depth across repeated noise draws. If it does not, the objective is not trustworthy, and we will fit the background discount $\gamma$ from data.
For real data we cannot re-pose the sensor after the fact. Instead, we will select views from the captured pool, if it contains more views than the reconstruction uses. We will emulate shorter exposures by binomially thinning recorded photons, which is statistically exact for Poisson streams.
Baselines
The naive reference is a uniform orbit with equal exposure per view, spending exactly the same total photon budget. It is run at 2, 3, 5 and 8 views. This is the configuration family Transientangelo evaluates, and we will run it through its released reconstruction pipeline. Beating it means lower Chamfer distance at matched photons across three or more noise seeds.
The strongest existing reconstructor in this regime is Transientangelo itself. The file reports only a relative figure: Chamfer distance about 5 times lower than TransientNeRF on the simulated Food and Cinema scenes. No reported absolute Chamfer result was found on Lacuna for Transientangelo, so we will reproduce its fixed-view numbers ourselves as the anchor.
The strongest competing planner is a curvature-aware next-best-view in the style of Path Matters. It is ported to lidar by scoring poses with the curvature of the current signed distance function, with the same B-spline smoothness penalty, the same budget and the same reconstructor. It receives uniform exposure, and also a variant with curvature-proportional exposure, so that the comparison isolates the choice of objective. For RGB reference, Path Matters reports PSNR 20.65 on DTU with 3 views and 27.05 on Mip-NeRF 360 with 12 views. These numbers are not comparable to lidar Chamfer distance and serve only to show that the method is strong in its own regime.
The third baseline is an uncertainty planner that moves to the pose of highest ensemble disagreement, using five reconstruction seeds. It also uses Transientangelo's weight-variance signal, and it has the same budget and tuning time.
Three ablations isolate our ingredients. The first is a pose-only variant with uniform exposure. The second is an exposure-only variant on uniform poses. The third replaces our objective with ensemble variance multiplied by the predicted photon yield, which tests whether the yield term carries the benefit.
Why not just…?
Why not just use a uniform orbit and dwell longer? At high flux this may be nearly as good, and we expect the gap to shrink toward the 6000 photons-per-pixel end. At matched total photons the question is allocation, not quantity. A uniform orbit spends equal time on grazing and occluded views, whose $\cos^3\alpha / r^2$ yield is small. The radiotherapy result in Beam orientation optimization, where 6 optimized beams beat 11 equiangular ones, suggests that orientation cannot be bought back with intensity. Our photon-count sweep will show where this stops being true.
Why not just run the curvature-aware planner from Path Matters? That planner assumes clean depth. At 10 photons per pixel, curvature of the current estimate is dominated by spurious ripples. We will test this directly. Using the ground-truth mesh, we will measure what fraction of the curvature planner's chosen views target true high-curvature regions versus noise-induced ones, as a function of photon count. If curvature planning matches ours at low counts, the noise-chasing concern is wrong and we will report that.
Why not just go where the model is most uncertain? This is the most serious threat. Ensemble disagreement does capture where the shape is unknown. It does not know that a candidate view will return few photons from a given patch, because yield depends on angle, range and reflectivity. Ensemble disagreement also gives no rule for splitting exposure time. The variance-times-yield ablation settles this. If it matches our planner, then the contribution is the photon-yield model, which can be added to any uncertainty planner, and we will frame the result that way.
Why not just denoise first and then plan with standard tools? In Photon-Starved Scene Inference, denoise-then-infer reached 27.21% accuracy at 0.11 photons per pixel, against 42.37% when noise was modeled directly. Transientangelo similarly gains by fitting raw histograms. We therefore keep counts raw throughout.
Proposed Experiments
The primary experiment sweeps total photon budgets from 6000 down to 10 photons per pixel equivalent, on the simulated Food and Cinema scenes and the other simulated scenes. It reports Chamfer distance, normal consistency and per-patch depth error for every baseline and ablation at matched budgets.
The decisive check has two parts. The planned-over-uniform advantage should grow monotonically as photons decrease. The curvature planner's share of views aimed at true geometric features should fall with photon count while ours does not.
A calibration check plots predicted against empirical depth standard deviation per patch. A slope near one validates the objective.
On real captures, exposure allocation via photon thinning, and pose selection from the captured pool where available, test whether the simulated gains survive real background and timing jitter. Collision and travel constraints are reported as a secondary cost-quality curve rather than a headline claim.
References
- Transientangelo: Few-Viewpoint Surface Reconstruction Using Single-Photon Lidar
- Path Matters: Unveiling Geometric Implicit Bias via Curvature-Aware Sparse View Optimization
- Photon-Starved Scene Inference using Single Photon Cameras
- Uncertainty-driven view planning for underwater inspection
- Beam orientation optimization for intensity modulated radiation therapy using adaptivel2,1-minimization
- Embodied Crowd Counting
- Learning Virtual View Selection for 3D Scene Semantic Segmentation