Contour Context Selection for Object Detection: A Set-to-Set Contour Matching Approach

Abstract

We introduce a shape detection framework called Contour Context Selection for detecting objects in cluttered images using only one exemplar. Shape based detection is invariant to changes of object appearance, and can reason with geometrical abstraction of the object. Our approach uses salient contours as integral tokens for shape matching. We seek a maximal, holistic matching of shapes, which checks shape features froma large spatial extent, as well as long-range contextual relationships among object parts. This amounts to finding the correct figure/ ground contour labeling, and optimal correspondences between control points on/around contours. This removes accidental alignments and does not hallucinate objects in background clutter, without negative training examples. We formulate this task as a set-to-set contour matching problem. Naive methods would require searching over 'exponentially' many figure/ground contour labelings. We simplify this task by encoding the shape descriptor algebraically in a linear form of contour figure/ground variables. This allows us to use the reliable optimization technique of Linear Programming. We demonstrate our approach on the challenging task of detecting bottles, swans and other objects in cluttered images. Abstract. We introduce a shape detection framework called Contour Context Selection for detecting objects in cluttered images using only one exemplar. Shape based detection is invariant to changes of object appearance, and can reason with geometrical abstraction of the object. Our approach uses salient contours as integral tokens for shape matching. We seek a maximal, holistic matching of shapes, which checks shape features from a large spatial extent, as well as long-range contextual relationships among object parts. This amounts to finding the correct figure/ground contour labeling, and optimal correspondences between control points on/around contours. This removes accidental alignments and does not hallucinate objects in background clutter, without negative training examples. We formulate this task as a set-to-set contour matching problem. Naive methods would require searching over 'exponentially' many figure/ground contour labelings. We simplify this task by encoding the shape descriptor algebraically in a linear form of contour figure/ground variables. This allows us to use the reliable optimization technique of Linear Programming. We demonstrate our approach on the challenging task of detecting bottles, swans and other objects in cluttered images.

Weakly Supervised Shape Based Object Detection with Particle Filter

“…3. According to the curves, we are better than Lu et al [18], Ommer and Malik [20], Ferrari et al [7,6] and perform equally well as Zhu et al [27]. The performance of the proposed method illustrates its ability to cope with substantial nonrigid deformations, which are present in the class Giraffes.…”

“…3 reports precision-recall (P/R) curve and detection rate vs false positive per image (DR/FPPI) curve for the class Giraffes in ETHZ dataset. In P/R, we compare to Lu et al [18], Zhu et al [27], Ommer and Malik [20] and Ferrari et al [7], whose results are quoted from [18]. In DR/FPPI, as Ferrari et al [7,6], Ommer and Malik [20] and Lu et al [18] provide their results, we compare to them.…”

“…Ommer and Malik [20] propose a novel Hough voting strategy to overcome the problem of scales. Zhu et al [27] treat the detection as a set-to-set matching problem between segments. They simplify the problem into linear programming to reduce the complexity.…”

“…They are learned automatically with partially-supervised learning. While many state-of-the-art approaches construct part models manually [18,27], we limit manual labeling to a single contour. In our approach, only one silhouette is manually decomposed into visual parts in advance.…”

From a Set of Shapes to Object Discovery

“…Existing work that uses only shape cues for recognition in real-world images requires either a manually specified shape template [4,5], or manually segmented training images to learn the object shape [6]. Also, all previous work on unsupervised object-category discovery exploits the photometric properties of segments [7,8], textured patches [9], and patches along image contours [10].…”