Reproduce “Sketched RT3D: How to reconstruct billions of photons per second” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Julián Tachella, Michael P. Sheehan, Mike E. Davies
SRT3D combines compact online photon sketches with RT3D’s spatially regularized point-cloud reconstruction to recover 3D lidar scenes without storing every photon arrival or using fine temporal histograms. Using only 5–10 complex sketch components per pixel, it achieves reconstructions comparable to full-data RT3D while reducing processing and memory growth as photon counts increase, reaching roughly 14 frames per second for a 705×705 array on an RTX 3070.
Key insights
- Spatial regularization is most valuable in low-photon and low-SBR regimes: SRT3D and RT3D recover coherent surfaces where pixelwise estimators produce incomplete or unstable results.
- A sketch size of 5 is often adequate in favorable conditions, while 10 components generally match the reconstruction quality of the full temporal data on harder examples.
- Because the sketch size is fixed, memory use and core reconstruction cost are largely decoupled from the number of detected photons per pixel.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.