Edge Detection

You can find more details (including examples, pre-calculated ETHZ results and code) on our team homepage at Virtual Habitat: http://vh.icg.tugraz.at

Many of the recently popular object recognition methods use shape as underlying representation, since for a wide variety of object categories shape matters more than local appearance. Shape is in general a quite powerful feature since an object contour is invariant to extreme lighting conditions and large variations in texture or color and additionally allows delineating accurate object outlines.

Most of these methods implicitly require stable, connected and labeled edges in the image as input to the algorithm, which has the main purpose of reducing the data amount while retaining the important information about the image content. Most approaches rely on a post-processed Canny or Berkeley (and its recent extensions) edge detection results, which are considered as state of the art in this field.

We have proposed a novel edge detection method for the purpose of object localization in images, which extends purely local edge detectors by additionally analyzing mid-level cues, i.,e.~regions that support the local gradient magnitude are analyzed to extract the most stable edges. Our method is based on analyzing a hierarchical data structure denoted as component tree where connected regions, which are separated by high gradient magnitudes along their boundaries, are linked in a tree structure. Finding the most stable edges in the tree, by considering shape similarity of the region contours, removes noise and clutter and preserves the important edges for detection. Furthermore, our method automatically labels all obtained edges during calculation. Therefore, no further post-processing is required and the results can be directly passed to any of the available shape based object localization methods.

We evaluated our method on two well known object recognition data sets: ETHZ and Weizmann horses, demonstrating improved results in low computation time.

Following figure shows exemplary results for the well-known ETHZ recognition dataset, where default Canny edge detection results are shown in red and results of our novel edge detector are shown in green.