Search for: All records

In general, conventional computer vision techniques suffer from an inability to detect hidden surface contacts due to line-of-sight visibility problems. Rather than fitting models to scene objects and estimating inter-object gaps, our approach is to leverage the fact that light passing between and reflecting off the surfaces can offer valuable information as it alters the appearance of nearby surfaces. For a proof of concept demonstration, we employed a machine learning approach to classifying adjacent surface imagery to estimate hidden surface distances and contact locations in a controlled setting under ambient lighting conditions. Our proofof- concept results demonstrate relatively high accuracy for the estimation of gap size and the detection of contact between hidden surfaces. We envision such measures could someday provide complementary information to be combined with traditional visible-surface methods, to obtain more precise and robust estimates of hidden surface relationships.