Enhancing Visual Object Detection with Semantic Tagging
Shiau Hong Lim and Peter Auer
Incorporating additional information into the learning system can often be the key to improving its generalization performance, especially in challenging tasks where training data is limited. However, even when extra information is readily available or can be easily obtained, it is often unclear how they can be utilized in the system. We present a simple method for incorporating extra information in the form of semantic tags through a probabilistic Hough transform framework. Semantic tagging allows features that are different in their native representations to be linked through their higher-level ”meaning” and therefore enhances the capability of a learner to generalize. We evaluate our method on both synthetic and real-world visual object detection tasks, and show that systems enriched with semantic information achieve state-of-the-art performance using fewer training examples.