Filtering, segmentation and depth (Q684557)

This volume reports new developments in computer vision research. The authors manage to make some progress at several levels of visual reconstruction by working in a restricted world of single, stationary pictures of interposed objects with uniform non-textured surfaces where shadows are not present. In the attempt to develop a computer algorithm that finds ``the 2.1D sketch of an image'' (i.e., performing image segmentation along with computing the relative depth of objects in an image), the authors suggest a practical way of incorporating occlusion information into the task of finding object outlines. Accordingly, the model proposed in this book determines the nearness relations of the objects from overlap analysis. The book is organized in sections that describe the algorithm stages. Chapter 2 is concerned with a new edge-enhancing filter, in fact a nonlinear filter that localizes edges while preserving corners and \(T\)- junctions, the crucial cues to occlusion. Chapter 3 covers the contour- detection problem by providing a fast curve smoothing algorithm. Chapter 4 is devoted to the computation of continuations of disrupted edges. Chapter 5 treats the final stage that seeks to find the optimal segmentation and the depth relations between regions. Results are analyzed for a number of images that yield correct results and others that yield wrong interpretations. Finally, Chapter 6 reviews the strengths and limitations of the algorithm and discusses areas of potential improvement. The text is supplemented by appendices giving the filter algorithm and the contour tracing algorithm in pseudo-\(C\) code. The program code takes up more than one third of the book (51 pages). The reviewer was rather disappointed to find out several errors (e.g., an infinite FOR loop on page 104 where a BREAK condition is missing, the function TraceContour is defined on page 104 as having 4 arguments and called on page 100 with only 3, the functions GrowContours and DeleteStray are called with real arguments although they are defined with integer ones). It is worth mentioning that, by making the actual programs used in this book available via internet, the authors serve the computer vision community in the efforts to have serious long-term progress. The possibility of sharing computer code is essential for the researchers who, in this way, are enabled to achieve future results in short time by combining the best parts of the previous ones.