Part based object detection with a flexible context constraint

This work describes an object detection system which integrates flexible spatial context constraints to improve detection performance. It allows spatial and scale deformation of the object relative to its context. The contextual model extends an existing deformable parts model and is trained on partially labeled data using a latent SVM. The approach can be applied to any object detection problem where the object class always exists in one typical image context, but the context can appear independently. A new scoring method is used to model the asymmetric relationship between object and context. Furthermore, the system enables the use of contextual non-maximum suppression, a context sensitive way to discard redundant detections. Trained on our combined dataset of dresses and persons, the system achieves a significant improvement in detection performance when compared with basic deformable parts models.