BayesToF: multiresolution denoising of indirect time-of-flight distance maps.

Area of Science:

• Computer Vision
• Computational Imaging
• Sensor Technology

Background:

• Indirect time-of-flight (IToF) cameras reconstruct depth but suffer from noise (photon, thermal, ambient).
• Existing denoising methods struggle with IToF's nonlinearities and simplified noise models, limiting real-world accuracy.
• Millimeter-level distance detail is often lost due to noise in raw IToF measurements.

Purpose of the Study:

• To develop a robust, physics-aware denoising framework for IToF cameras.
• To address the challenge of modeling the complex noise in IToF measurements.
• To improve the accuracy and reliability of IToF-based distance reconstruction.

Main Methods:

• Derived a structure-aware Bayesian denoising framework for IToF.
• Utilized a multinomial-Poisson process for tractable joint distribution modeling.
• Employed a logarithmic transformation and Gaussian-scale-mixture prior for linearization.
• Developed a multi-resolution hierarchical Bayesian minimum-mean-square-error (MMSE) estimator.

Main Results:

• The BayesToF method achieves robust noise suppression without network training.
• It automatically adapts to varying noise levels and spatial frequencies.
• Experiments show superior accuracy and generalization compared to state-of-the-art methods on public and new datasets.

Conclusions:

• BayesToF provides a significant advancement in IToF denoising by leveraging physics-based modeling.
• The framework enables high-quality distance reconstruction with improved robustness.
• This approach overcomes limitations of previous methods relying on simplified noise models.