sher@rochester.UUCP (09/07/85)
From: David Sher <sher> I accidently sent the beginning of this article before it was finished so if you see an incomplete version of this I apologize. Anyway, I recently presented a paper to the Workshop on Probability and Uncertainty in AI on detecting boundary points (like edges but unoriented) from small windows based on probabilistic scene models. I encountered some criticism with regard to the "realism" of my scene models. Here is the scene model I used: 1. Regions (the things the boundaries are about) have no internal variation in graylevel (I am using black and white currently). 2. The noise has no known spacial structure and the probability of observing a gray-level at a point with a specified gray-level is known. 3. A small (3x3) window only intersects one boundary at a time or no boundary at all. 4. Only the (3x3) window is relevant to the probability of a boundary. (This is the meaning of restricting the operator to a window.) 5. The misregistration is a sufficent problem that the spactial structure within the small (3x3) window is not significant. Using models of this form I directly derived certain near optimal boundary detectors and analyzed some other simple ones. The reason I chose models of this form is the optimal operators vary in complexity with the complexity of the model and their computational complexity varies exponentially (intuition not mathematics) with the complexity of the model. This puts a large cost on naive analysises with complex models. Also this kind of model is close to the models commonly used by researchers in this area (such as Canny). I am now considerring more complex models but am not sure which way to go. Some possible more general models would: - Replace assumption 5 with a description of the possible spacial structures of small windows on and off boundaries. - Add blur type noise to assumption 2 (I am not sure exactly the best way to do this) - Replace assumption 1 with an assumption about the reflectance (or possible reflectances) of the objects and the sun position Any suggestions? By the way I have seen: %A Leonard P. Wesley %A Allen R. Hanson %T The Use of an Evidential-Based Model for Representing Knowledge and Reasoning about Images in the Visions System %K VISION DEMPSTER SHAFER EVIDENCE PROBABILITY THEORY %J PAMI %D Sept 1982 %V 4 %N 5 %P 14-25 %I IEEE %X Describes how Dempster-Shafer evidence theory is used for interpretation of a segmented image. Takes segemented image with statistics about each segment and binary relationships between segments and applies a known probability structure to it. %A David Sher %T Developing and Analyzing Boundary Detection Operators Using Probabilistic Models %K VISION PROBABILITY STATISTICS BAYES %J Workshop on Probability and Uncertainty in Artificial Intelligence %D August 1985 %I ACM,RCA %X Describes how evidence from boundary detectors can be combined if the boundary detectors take on a certain form. This form is to generate the likelihood of a boundary rather than the probability. Demonstrates how boundary detectors of this form can be constructed directly from a probabilistic model of the image and its boundaries. Also demonstrates how an established operator, the gradient, be converted into an operator of this form. Discusses what this tells us about the gradient. %A Hsien-Che Lee %A King-Sun Fu %T Generating Object Descriptions for Model Retrieval %D Sept 1983 %K VISION SURFACE 3D MODEL %J PAMI %V 5 %N 5 %P 462-471 %I IEEE %X Describes a system to return 3-D information starting with grey level image. First pass segments the image according to local grey level information such as gradients. Next uses this to select a target area. Then does edge detection & linking and forms chain codes of edges. Uses the result to extract regions with a slight perference to over fragmentation which will be reversed by 3D information. Uses shape of regions and regularity constraints like skew symmetry to derive 3D information. Picks region whose position indicates least slant and uses that to further constrain the image. Like Waltz filtering but on plane intersections rather than on junctions, thus more robust. %A John Francis Canny %T Finding Edges and Lines in Images %K VISION EDGE LINE %D June 1983 %I MIT Artificial Intelligence Laboratory %R 720 %X I quote from the Conclusion: ... %A Vishvjit S. Nalwa %T On Detecting Edges %K VISION EDGE %J Proceedings: Image Understanding Workshop %D October 1984 %P 157-164 %I Image Processing Techniques Office, Defense Advanced Research Projects Agency %X Describes a very good edge detector. It tries to fit a serries of surfaces over a window. If a plane in intensity space fits the window well then it isn't an edge. If a model of a step edge fits better than a quadratic surface then it is an edge. Determine the depth of the edge from the surface that was fit to the window. %A Thomas O. Binford %T Inferring Surfgaces from Images %K VISION MODEL SURFACE EDGE %J Artificial Intelligence %D August 1981 %V 17 %N 1-3 %P 205-244 %I North-Holland Publishing Company %C Amsterdam %X Describes a model of image description. Gives a set of criterion for interpretation of various kinds of 3-d edges,lines and discontinuities within them. Describes some criterion for edge detectors and a hack which binford particularly likes. %A Larry S. Davis %A Azriel Rosenfeld %A Steven W. Zucker %T General Purpose Models: Expectations about the Unexpected %K VISION MODEL %D January 1975 %I University of Maryland %R TR-347 %X Shows a need for models that do not describe specific situations. Describes some of the features such models might have. Describes when such models might be necessary and when they are unnecessary. General purpose models are applicable even when we have little or no a priori knowledge about the class of scenes that is to be analyzed. This work is notably vague. %A Allen R. Hanson %A Edward M. Riseman %T VISIONS: A Computer System for Interpretting Scenes %K VISION VISIONS HIGH LEVEL MODEL %D 1978 %P 303-334 %I Academic Press %B Computer Vision Systems %E Allen R. Hanson %E Edward M. Riseman %C New York %C San Francisco %C London %X Description of high level part of the VISIONS system. Assumes segmented images. works largely by matching segments with similar parts of models. Matching strategy determined flexibly. Strategy optimized to model being matched. Use of model dependent information emphasized. Use of knowledge of 3D constraints aeffects also. %A Dana H. Ballard %A Chris M. Brown %A Jerome A. Feldman %T An Approach to Knowledge-DIerected Image Analysis %K VISION HIGH LEVEL MODELS %D 1978 %P 271-282 %I Academic Press %B Computer Vision Systems %E Allen R. Hanson %E Edward M. Riseman %C New York %C San Francisco %C London %X Describes how a vision system might work. Introduces the concept of the sketch map between low level segmentation and high level model. The system described is query directed. The system seems to be bottom up to sketch map and top down to sketch map. -David Sher sher@rochester seismo!rochester!sher -- -David Sher sher@rochester seismo!rochester!sher