FirstCrossConvex Single-Photon 3D Reconstruction from Sparse Views

Chi-Wei Lee, Cheng-Han Chiang, Yi-Chun Chang, An-Kai Chen, Po-Chih Kuo

National Tsing Hua University, Hsinchu, Taiwan

FirstCross reconstructs 3D shape from very few photons. Transientangelo appears at full flux, at 10 PPP and at 1 PPP with adapted settings on photons pooled over 4 × 4 pixels.

Abstract

Single-photon lidar records the arrival time of each returning photon, enabling 3D reconstruction from few views and scarce photons. However, leading neural methods such as Transientangelo require over an hour of non-convex per-scene optimization, and their accuracy degrades as photon flux decreases. We propose FirstCross, a network-free approach that instead solves a short sequence of convex problems on a voxel grid. For a solid object, the Poisson likelihood of a ray with given signal and background rates depends on the scene solely through its first crossing, the point where the ray first enters the object. Decomposing this likelihood yields two tractable terms. The first penalizes surfaces in front of detected photons and admits a tight convex closure, namely volume rendering with a running-maximum transmittance in place of NeRF's product transmittance. The second accounts for photons that no surface explains and admits a convex upper bound once the object is assumed to be thick behind each detected surface point. Compared with Transientangelo under its own real-data protocol, FirstCross achieves lower depth error across every test scene, view count, and flux level. With three views and 100 times fewer photons than at the protocol's lowest flux, our mean object depth error stays under half of Transientangelo's with all photons. Given our detected surface points, the best classical fusion has 1.3 to 1.8 times our three-view error across flux levels. On real room-scale captures with 1% of their photons, TransientNeRF has 2.8 to 7.7 times our error. FirstCross lets few-view single-photon scanners cut exposure to 1% and reconstruct an object in about a minute on one GPU.

Method

Overview of FirstCross on real data: photons of one view, the first crossing of a ray, the convex iterations and the resulting shape
A ray’s Poisson likelihood splits into two terms. One penalizes a surface in front of detected photons, and its tight convex closure is volume rendering with a running-maximum transmittance.

Results

Photon levels

Baskets, 2 views.

MethodFull flux10 PPP1 PPP0.1 PPP
Ground truth Ground truth mesh of the Baskets scene
FirstCross FirstCross novel view at full flux0.90 mm FirstCross novel view at 10 PPP0.89 mm FirstCross novel view at 1 PPP0.95 mm FirstCross novel view at 0.1 PPP1.17 mm
Transientangelo Transientangelo novel view at full flux2.60 mm Transientangelo novel view at 10 PPP4.94 mm
None below 10 PPP
None below 10 PPP
Mean object depth MAE over the three-view test scenes is printed under each tile. PPP is the mean photon count of an occupied pixel.

Mean depth error with the turntable back wall, in millimetres, at each photon level, three views

PPPFull300150501010.1
FirstCross2.972.962.972.952.973.033.25
Transientangelo4.704.484.78none7.04nonenone
Our mean depth error stays between 2.95 and 2.97 mm down to 10 PPP. Transientangelo's goes from 4.70 to 7.04 mm over the same range.

Meshes

Chef at 1 PPP, 3 views.

Transientangelo mesh of chef at 1 PPPFirstCross mesh of chef at 1 PPP
With its photons pooled as in ours, the adapted Transientangelo recovers chef and baskets at 1 PPP with 7.6 and 1.9 times our depth error.

Training

Chef at 1 PPP, 3 views.

Both methods build chef from the same photons while the camera turns once around it. Transientangelo runs from step 100 to step 150,000, over an hour of training, and ours from round 1 to round 12, about a minute. Orange marks surface that is not in the true shape and blue marks true surface that is missing, at a 2 cm tolerance.

Transients

Ours and Transientangelo are compared with the full-flux measurement, each trained at full flux in the upper row and at 10 PPP in the lower row. All panels share one log scale.

Rooms

Ground truthFull flux1% flux
Depth of the test rooms at full and 1% flux for TransientNeRF and FirstCross TransientNeRFOurs TransientNeRFOurs TransientNeRFOurs
On the real rooms of SP-TransientBench with three views, TransientNeRF's depth at 1% flux becomes noisy with spurious surfaces in every room, while ours stays near its full-flux depth. Volumetric (TSDF) fusion of our detected returns still outperforms TransientNeRF at 1% flux.

BibTeX

@article{lee2026firstcross,
  title   = {FirstCross: Convex Single-Photon 3D Reconstruction
             from Sparse Views},
  author  = {Lee, Chi-Wei and Chiang, Cheng-Han and Chang, Yi-Chun
             and Chen, An-Kai and Kuo, Po-Chih},
  journal = {arXiv preprint arXiv:2611.26321},
  year    = {2026}
}