Unsupervised Domain Adaptation of Deep Networks for ToF Depth Refinement.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Depth maps from Time-of-Flight (ToF) cameras suffer from noise and multi-path interference, limiting accuracy.
• Training deep networks for depth denoising typically requires expensive ground truth data from real-world acquisitions.
• Domain shift between synthetic and real data hinders the performance of networks trained solely on synthetic data.

Purpose of the Study:

• To develop unsupervised domain adaptation techniques for refining ToF depth maps.
• To enable deep learning models trained on synthetic data to generalize to real-world ToF data without requiring ground truth.
• To address the challenges of noise and multi-path interference in ToF depth sensing.

Main Methods:

• Proposed three unsupervised domain adaptation approaches: input-level (domain translation), feature-level (feature alignment), and output-level (adversarial loss).
• Utilized domain translation networks to adapt synthetic ToF data representation.
• Employed adversarial loss with a discriminator to train the denoiser on unlabeled real data.

Main Results:

• The proposed methods demonstrated superior denoising performance compared to state-of-the-art techniques.
• Unsupervised domain adaptation effectively bridged the gap between synthetic and real ToF depth data.
• Achieved significant improvements in depth map accuracy for ToF cameras.

Conclusions:

• Unsupervised domain adaptation is a viable and effective strategy for training depth denoising networks on real-world ToF data.
• The presented input, feature, and output level adaptation methods offer robust solutions for ToF depth refinement.
• These advancements pave the way for more accurate and accessible ToF depth sensing applications.