DeepToF: Off-the-Shelf Real-Time Correction of Multipath Interference in Time-of-Flight Imaging

TL;DR

DeepToF introduces a real-time, hardware-free neural network-based method to correct multipath interference in ToF cameras, significantly improving depth accuracy without additional hardware or costly computations.

Contribution

The paper presents a novel autoencoder-based approach for MPI correction that requires no camera modifications and operates in 10 milliseconds per frame.

Findings

Effective MPI correction on real and synthetic data

No hardware modifications needed

Real-time processing capability

Abstract

Time-of-flight (ToF) imaging has become a widespread technique for depth estimation, allowing affordable off-the-shelf cameras to provide depth maps in real time. However, multipath interference (MPI) resulting from indirect illumination significantly degrades the captured depth. Most previous works have tried to solve this problem by means of complex hardware modifications or costly computations. In this work we avoid these approaches, and propose a new technique that corrects errors in depth caused by MPI that requires no camera modifications, and corrects depth in just 10 milliseconds per frame. By observing that most MPI information can be expressed as a function of the captured depth, we pose MPI removal as a convolutional approach, and model it using a convolutional neural network. In particular, given that the input and output data present similar structure, we base our network in an autoencoder, which we train in two stages: first, we use the encoder (convolution filters) to learn a suitable basis to represent corrupted range images; then, we train the decoder (deconvolution filters) to correct depth from the learned basis from synthetically generated scenes. This approach allows us to tackle the lack of reference data, by using a large-scale captured training set with corrupted depth to train the encoder, and a smaller synthetic training set with ground truth depth to train the corrector stage of the network, which we generate by using a physically-based, time-resolved rendering. We demonstrate and validate our method on both synthetic and real complex scenarios, using an off-the-shelf ToF camera, and with only the captured incorrect depth as input.