Synthetic Data Generation for Demonstrating Noise Reduction in Facial Depth Imaging
Keywords: Depth map, Denoising, Synthetic data
TL;DR: In most modern papers, the denoising of depth data is only a small section of the paper and compared to traditional denoising only limited evaluation values given.
Abstract: Deep learning has achieved remarkable success in image denoising, especially for photographic data. However, advancements in denoising depth images, particularly those captured by Time-of-Flight (ToF) sensors, have been limited due to the scarcity of clean ground truth data. This study introduces a method for generating synthetic facial depth data that closely emulates the noise characteristics of ToF sensors, facilitating the creation of paired clean and noisy datasets for supervised learning. We evaluate state-of-the-art convolutional neural networks (CNNs) on these synthetic datasets to assess their denoising performance. The findings demonstrate that synthetic datasets can effectively train depth-denoising models, thus enhancing the quality of facial depth maps in practical applications. Our results suggest that using synthetic data to create realistic, noisy, and clean datasets can highlight denoising performance through advanced techniques.
Submission Number: 69
Loading