Gradient-guided filtering of depth maps using deep neural networks

Image filtering has long been an area of interest in computer vision applications. It may be used in applications where noise is prominent or when certain features need to be enhanced. While single-image filtering techniques are well-established in literature, the introduction of additional information can further improve the quality of filtering. Guided filtering allows for the use of additional signals to enhance the filtering process. However, many of these techniques operate on natural images and are not suited for certain classes of images such as depth maps. In this work, we propose a filter that is specifically tuned to operate on noisy depth maps. To guide the filtering process, known image gradients are inputted into the system. Given the complex nature of this input, a five-layer neural network built using stacked denoising autoencoders was used to implement a black-box filter. Testing with the proposed system shows the benefits of using the deep network for depth map filtering.