Using Partial Edge Contour Matches for Efficient Object Category Localization

Object detection based on scale-invariant partial shape matching

“…But their performance decreases seriously when local shape distortion and occlusion occur. Partial shape descriptors [6,7] can somewhat overcome this issue for their capability of handling occlusion problem in the complex scene and allowing partial matching.…”

From partial shape matching through local deformation to robust global shape similarity for object detection

“…For purpose of improving detection and score ranking, a sophisticated training process is designed in which latent SVM is used to guarantee the many-to-one score is tuned discriminatively. Besides of literature mentioned above, edge information is also utilized in [19,15,3].…”

“…(10) Based on these local observations, the most popular method to form object location hypothesis is using Hough voting, such as in [19]: local maxima of S(i, j, k) for certain fragment e k are identified, and corresponding fragment correspondences are used to estimate object location by Hough voting. However, Hough voting seems not to be an optimal choice here.…”

“…This may result in missing the target object in the image or locating the object inaccurately, e.g., if the entire yellow fragment is assigned to the swam, the detected bounding box will be larger than the ground truth. To our best knowledge, only the approach in [19] explicitly addresses problem (3) by introducing an efficient partial matching schema based on integral image [22].…”

“…However, the final detection evaluation in [19] is appearance based (SVM on HOG features), which demonstrates weakness in the discriminative power of their partial matching schema. There are at least two main reasons for this, one is the selection of the best matching fragments in the integral image framework and the other is simply weak discriminative power of their shape descriptor, which is only angle based.…”