Powered by OpenAIRE graph
Found an issue? Give us feedback
image/svg+xml art designer at PLoS, modified by Wikipedia users Nina, Beao, JakobVoss, and AnonMoos Open Access logo, converted into svg, designed by PLoS. This version with transparent background. http://commons.wikimedia.org/wiki/File:Open_Access_logo_PLoS_white.svg art designer at PLoS, modified by Wikipedia users Nina, Beao, JakobVoss, and AnonMoos http://www.plos.org/ ZENODOarrow_drop_down
image/svg+xml art designer at PLoS, modified by Wikipedia users Nina, Beao, JakobVoss, and AnonMoos Open Access logo, converted into svg, designed by PLoS. This version with transparent background. http://commons.wikimedia.org/wiki/File:Open_Access_logo_PLoS_white.svg art designer at PLoS, modified by Wikipedia users Nina, Beao, JakobVoss, and AnonMoos http://www.plos.org/
ZENODO
Preprint
Data sources: ZENODO
addClaim

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

Authors: Lin, Po-Ting;

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

Abstract

Path-traced synthetic stereo data underlie a large fraction of moderndisparity-estimation training pipelines. We report a previouslyunrecognised property of such data: while the Monte~Carlo (MC) noisestreams of the two cameras are statistically independent, the underlying\emph{variance fields}---deterministic per-pixel functions of therendering integrand---are highly correlated once aligned by theground-truth disparity warp. Across 20 scenes rendered with Mitsuba~3,the warped Pearson correlation reaches $\rho{=}0.754{\pm}0.016$ across20 scenes at $\mathrm{SPP}{=}512$, and on a representative sceneremains essentially invariant ($\rho{=}0.778{\pm}0.001$) over a$16\times$ range of samples per pixel. The effect is strongest inLambertian regions ($\rho{\approx}0.78$) and substantially weaker inglass ($\rho{\approx}0.30$), as predicted by an integrand decompositioninto view-independent and view-dependent components. A residual-shuffle intervention that breaksthe cross-view alignment while preserving the clean image collapsesthe GT cost margin by $33\%$ on non-glass and degrades variance-basedwinner-take-all accuracy on glass by $4.3\times$, confirming thestructure functions as a matching cue. This signal is unique toMC-rendered data and constitutes a candidate sim-to-real shortcut whoseimpact on trained networks remains to be quantified.

Powered by OpenAIRE graph
Found an issue? Give us feedback