Modeling hierarchical and heterogeneous feature representation with conditional random field for visual object detection

We propose a novel flexible and hierarchical object representation using heterogeneous feature descriptors for detection of visual objects in real-world scenarios. Our representation is built on a Conditional Random Field (CRF) model that is able to aggregate local, semi-local and global features in one consistent framework. To improve the discriminative power of our model, we incorporate SVM classifiers into the CRF to learn discriminative unary classifiers for different object parts. Besides parameter learning of unary classifiers, a topology learning that captures the underlying geometrical structure of the target object class also boosts the performance of our model. Evaluation results on both simple UIUC single-scale car dataset and the challenging PASCAL VOC 2007 dataset verify that our model is flexible enough for a wide variety of object classes and robust to appearance variations caused by pose changes, articulation and partial occlusion.