Depth-assisted edge detection via layered scale-based smoothing

This paper deals with the subject of using depth data to aid the process of edge detection. Edge detection and stereo vision have had a long and abiding relationship, with several implementations of correspondence algorithms depending on edge detection techniques for robustness. The inverse however, has rarely been addressed. What if human vision utilizes stereoscopy (and hence, depth information) in order to aid the process of edge discovery, rather than the other way around? Our goal in this paper is to investigate this hypothesis and to establish whether the use of depth data can improve the performance of existing edge detection techniques. More specifically, we show how we used image distance estimates in order to improve the performance of the ubiquitous Gaussian smoothing filter in its role as a preprocessor for various edge-detection operators, such as the one developed by J. Canny (1986). Our approach ensures that the Gaussian smoothing filter is applied in such a way as to eliminate as much noise as possible, while preserving adequate detail in all parts of the image. This is accomplished by the `layering´ of the image into multiple separate sub-images and their subsequent independent smoothing. The layering process is guided by depth data, which is presumed to have been derived through stereoscopy or other means. The images are then re-composited and analyzed for edge information by an implementation of the Canny edge detector. Results are presented both with and without the application of our extended smoothing preprocessor for comparison