Deep Learning for Transient Image Reconstruction from ToF Data

Area of Science:

• Computer Vision
• Optical Sensing
• Machine Learning

Background:

• Multi-path interference (MPI) is a significant error in Time-of-Flight (ToF) cameras, causing depth overestimation due to light reflections.
• Existing methods struggle to accurately correct MPI, limiting the precision of ToF depth measurements.

Purpose of the Study:

• To develop a novel deep learning approach for correcting MPI in ToF cameras.
• To improve depth accuracy by distinguishing direct and global light components.
• To achieve state-of-the-art performance in MPI correction.

Main Methods:

• A deep learning model is proposed to estimate the scene's time-dependent impulse response.
• The model comprises a predictive block for encoding backscattering vectors and a fixed block for light response translation.
• The approach focuses on recovering depth images with reduced MPI from estimated light components.

Main Results:

• Experimental validation on real-world data demonstrates the effectiveness of the proposed method.
• The approach successfully reduces MPI, leading to more accurate depth estimations.
• The method achieves performance comparable to or exceeding current state-of-the-art techniques.

Conclusions:

• The novel deep learning approach effectively corrects multi-path interference in ToF cameras.
• Accurate estimation of direct and global light components is key to mitigating MPI.
• This work advances the precision of depth sensing in ToF imaging systems.