Zenodo (preprint) Preprint

Posted

Cross-View Variance Correlation in Path-Traced Stereo: A Hidden Shortcut in Synthetic Training Data

Po-Ting Lin 1

  1. 1 Independent Researcher
DOI
10.5281/zenodo.19892861
License
CC BY 4.0
Categories
Computer Vision · Stereo Matching · Synthetic Data

Path-traced synthetic stereo data reveal a previously unrecognized property: while the Monte Carlo noise streams between left and right cameras are independent, their variance fields show high correlation once aligned by ground-truth disparity. Across 20 Mitsuba 3 scenes, warped Pearson correlation reaches ρ = 0.754 ± 0.016 and remains invariant over a 16× sampling range. The effect is strongest in Lambertian regions and weaker in glass, which together constitute a learnable shortcut for stereo networks and a sim-to-real gap mechanism that is unique to rendered data and absent in real binocular sensors.

  • path tracing
  • Monte Carlo rendering
  • stereo matching
  • variance correlation
  • sim-to-real gap
  • learnable shortcut
  • Mitsuba 3
Loading PDF…
Cite as (BibTeX)
@misc{lin2026crossview,
  title = {Cross-View Variance Correlation in Path-Traced Stereo: A Hidden Shortcut in Synthetic Training Data},
  author = {Po-Ting Lin},
  year = {2026},
  howpublished = {Zenodo},
  doi = {10.5281/zenodo.19892861}
}

← All publications