Learning Coexistence Discriminative Features for Multi-Class Object Detection

Existing methods on object detection have the ability to learn the discriminative features of local regions for object recognition; however, the coexistence relation between the multi-class objects could also benefit recognition. In this paper, we propose to learn the coexistence discriminative features for multi-class object detection. Given an image with multiple class objects, the strong supervision of the region-based annotations are first used as the image-level label to learn the independent discriminative features for each class. Then, the coexistence relation is fused as coexistence feature based on the attention mechanism. By combining the independent discriminative features and coexistence feature, the classification performance of multi-class object proposals can be consistently improved. Experimental results prove that the proposed end-to-end network outperforms the state-of-the-art object detection approaches, and the learned discriminative features can effectively capture the coexistence relations to improve classification performance of multi-class objects in the object detection task.