Unsupervised Subcategories for Visual Recognition |
 |
|
Abstract: Many computer vision tasks are formulated as binary classification problems where all positive and negative examples are mapped to a common feature space and a single hyperplane separating them is estimated. Due to large intra-class variation in object appearance, object pose, and camera viewpoint, it is difficult to learn a single classifier that can achieve good performance. Recent advances have advocated splitting a category into smaller groups (subcategories) and training multiple classifiers per subcategory. While some approaches have partitioned the data using extra ground-truth annotations e.g., poselets, others have relied upon heuristics. There has been no clear consensus on how the partitioning has to be performed. In this work, we hypothesize that the common insight shared amongst the different methods is to partition the data into "visual subcategories" i.e., to group instances nearby in the feature space to yield visually homogeneous clusters. In addition to improvement in performance, we show that visual subcategories also facilitate the use of simpler representations and models, such as potentially alleviating the need for deformable parts. |
Publications"Context and Subcategories for Sliding Window Object Recognition" |