Learning Parts for Object Detection

Previous work on part-based models for object detection has concentrated on searching locally discriminative features representing objects based on notion of parts. There is little research on how to select parts effectively and what kind of parts could improve the object detections. This paper investigates the learning problem of object parts with weakly labeled data, and proposes an adaptive approach for part selection. Without part-level supervision, for each training example this approach first detects seed windows of parts using single-part classifiers and then localizes parts in local regions via the image-specific distribution. The selected parts, which contain discriminative and relevant features, are used to train global parameters. Addressing the partial object occlusions in training examples, a pruning strategy is introduced to reduce the proportion of noise parts during learning iterations. The experimental results on PASCAL VOC 2007 and 2010 datasets demonstrate that the proposed part learning method gets an improvement on object detections compared with three classical part models, and the pruning strategy can speed up the convergence rates of model learning.