Noise vs Feature: Probabilistic Denoising of Time-of-Flight Range Data

Advances in active 3D range sensors have enabled the recording of depth maps at video frame rates. Unfortunately, the captured depth data is often noticeably contaminated with noise. We present a series of statistical analyses and denoising techniques to improve the raw output of depth sensors. Unifying our investigations is the concept that in the presence of high sensor noise, the ability to distinguish effectively between noise corruption and real depth features is needed to reconstruct the original true geometry realistically.