1) PAPER TITLES 1. Evolving Boundary Detectors for Natural Images via Genetic Programming 2. Evolution of a Local Boundary Detector for Natural Images via Genetic Programming and Texture Cues 2) AUTHORS Ilan Kadar, ilankad@cs.bgu.ac.il Ohad Ben-Shahar, ben-shahar@cs.bgu.ac.il Moshe Sipper, sipper@cs.bgu.ac.il Physical address for all: Department of Computer Science, Ben-Gurion University, Beer-Sheva 84105, ISRAEL 3) CORRESPONDING AUTHOR Moshe Sipper 4) ABSTRACT Boundary detection constitutes a crucial step in many computer-vision tasks. We present a learning approach to automatically construct a high-performance local boundary detector for natural images. Our approach aims to use Genetic Programming (GP) as a learning framework for evolving computer programs that are evaluated against human-marked boundary maps, in order to accurately detect and localize boundaries in natural images. Our approach is unique in that it combines common early visual filter kernels, but at the same time it makes no assumptions about how these kernels interact and what constitutes a boundary in the first place, thus avoiding the need to make ad hoc intuitive definitions or constructions for either. By testing the evolved boundary detectors on a highly challenging benchmark set of natural images with associated human-marked boundaries, we show performance to be quantitatively competitive with existing computer-vision approaches. We present our best evolved boundary detector---GP/TG Detector---that outperforms most existing approaches. Moreover, we show that our best evolved detectors provide insights into the mechanisms underlying boundary detection in the human visual system. 5) CRITERIA (B) The result is equal to or better than a result that was accepted as a new scientific result at the time when it was published in a peer-reviewed scientific journal. (C) The result is equal to or better than a result that was placed into a database or archive of results maintained by an internationally recognized panel of scientific experts. (D) The result is publishable in its own right as a new scientific result, independent of the fact that the result was mechanically created. (E) The result is equal to or better than the most recent human-created solution to a long-standing problem for which there has been a succession of increasingly better human-created solutions. (F) The result is equal to or better than a result that was considered an achievement in its field at the time it was first discovered. (G) The result solves a problem of indisputable difficulty in its field. 6) WHY THE RESULT SATISFIES THE CRITERIA Why the result satisfies criterion (B) -------------------------------------- The literature on local boundary detection is vast and extensive. To name but a few publications (in top journals and conferences, e.g., PAMI, ICCV and CVPR): (*) J. F. Canny, A computational approach to edge detection, PAMI. (*) P. Perona and J. Malik, Detecting and localizing edges composed of steps, peaks and roofs, ICCV. (*) Stella X. Yu, Segmentation Induced by Scale Invariance, CVPR. (*) G. D. Joshi and J. Sivaswamy, A computational model for boundary detection, Lecture notes in computer science Our approach outperforms ALL the approaches mentioned above and many more on a highly challenging benchmark set of NATURAL images, which are much harder to deal with than computer-generated images. The result was accepted, among others, to ICPR, one of the top conferences in pattern recognition, i.e., the result was considered significant in the application field, outside the domain of evolutionary computation. Why the result satisfies criterion (C) -------------------------------------- We used the Berkeley Segmentation Dataset and Benchmark (BSDB), which contains 300 NATURAL images, each of which was manually segmented by human subjects. The dataset is divided into two independent sets of images: a training set of 200 images and a test set of 100 images. In order to ensure the integrity of the evaluations, only the images from the training set can be accessed during the learning process. This highly challenging benchmark is considered the de facto standard for measuring boundary detection performance. The obtained score of our approach ranks 2nd on BSDB among all local boundary detectors (see comment regarding 1st place below). Many datasets used in computer vision are small, contain simple images, and were created by computer-vision researchers. BSDB broke this tradition by using a wide variety of images (e.g., animals in natural scenes, people, manmade structures, and urban scenes) taken from the Corel dataset. In addition, the database contains several hundred images, orders of magnitude more than used by many computer-vision researchers to test their techniques. Why the result satisfies criteria (D,E,F) ----------------------------------------- Our best evolved operators outperform all the approaches that are based on convolving the image with common early vision filter kernels, including the oriented energy approach (also known as the "quadrature energy"), considered by many as an excellent model to detect boundaries in natural images. This suggests that our approach found the best-known operator in the search space of interest. The only better-performing operator on BSDB---brightness gradient combined with texture gradient---is completely outside the search space of interest. Brightness gradient is a complex operator, which requires tuning, and it is not inspired by any model of processing in the primate visual system, meaning that our approach found the BEST operator inspired by models of processing in the primate visual system. Moreover, we showed that our best evolved detectors provide insights into the mechanisms underlying boundary detection in the human visual system. Why the result satisfies criterion (G) -------------------------------------- Boundary detection is one of the best-known problems in computer vision, and has been vigorously addressed for the last 40 years. The performance of many high- level computer-vision tasks, such as segmentation and object recognition, is highly dependent upon the extracted boundary map of an image. 7) CITATIONS (*) I. Kadar, O. Ben-Shahar, M. Sipper. "Evolving Boundary Detectors for Natural Images via Genetic Programming." Proc. 19th International Conference on Pattern Recognition, Dec 2008, Tampa, Florida. ICPR 2008. (*) I. Kadar, O. Ben-Shahar, M. Sipper. "Evolution of a Local Boundary Detector for Natural Images via Genetic Programming and Texture Cues." 4th International Conference on Autonomous Robots and Agents. Feb 09. Wellington, New Zealand. ICARA 2009. (Note: Accepted but withdrawn due to lack of travel funding). (*) I. Kadar, O. Ben-Shahar, M. Sipper. "Evolution of a Local Boundary Detector for Natural Images via Genetic Programming and Texture Cues." Genetic and Evolutionary Computation Conference. July 2009. Montreal, Canada. GECCO 2009. 8) STATEMENT OF PRIZE DISTRIBUTION Any prize money, if any, is to be divided equally among the co-authors. 9) COMPARISON TO OTHER HUMAN-COMPETITIVE ENTRIES Boundary detection is a fundamental, important, and highly complex problem in computer vision. We have developed a learning framework, based on GP, for evolving boundary detectors. In many computer-vision papers researchers show their results on a VERY small number of images (i.e., less than 10, sometime less than 5) and point out why their results "look good." But we can never really gather from such studies how general the results are, given the dearth of images involved in the testing. (For example, "Synthesis of Interest Point Detectors Through Genetic Programming," GECCO 2006, which won a HUMIES award that year, uses only three (3) images for evaluation.) As opposed to such common and numerous studies, we used a well-known benchmark (BSDB), which is considered the de facto standard for measuring boundary-detection performance. The database is large and contains bona fide hard images. Experiments showed that our approach outperforms all-but-one computer-vision approaches. Moreover, we found the BEST operator inspired by models of processing in the primate visual system. (The only operator that "beat" us is far less interesting given its total unrelatedness to anything biological). Our results provide insights into the mechanisms underlying boundary detection in the human visual system in general. Therefore, our research has made a significant contribution from both a computational as well as a biological point of view.