Hierarchical multiple instance learning for object detection
MetadataShow full item record
Object detection remains one of the most important problems in computer vision. In the last decade, hierarchical representations have been proven powerful for object detection and recognition under general settings, achieving very low false positive rates. However, most of these approaches need large amounts of labeled data, which is often costly, time-consuming, and error prone. This problem can be alleviated using Multiple Instance Learning (MIL) which is a weakly supervised learning method that involves less human intervention yet still has been shown to accurately model class variation on a variety of problems. Motivated by the power of hierarchical representations and MIL, we propose Hierarchical Multiple Instance Learning, which is a variant of multiple component methods where each level of the hierarchy is considered as a component. MIL problems are sequentially posed and solved in a hierarchical manner to ultimately extract multi-component object models. We combine the hierarchical levels with a set support vector machines to define our final object detection models. We experiment and test our model on two challenging datasets: tumors in gastrointestinal capsule endoscopy (CE) and flowers in natural images. Our results clearly show the merit of combining hierarchical representations and MIL. We have chosen these two data sets to show the efficacy of the algorithm in real life object detection as well as in medical domain. Capsule endoscopy is a relatively new and almost unexplored field which requires different pathology detection and classification work. We have used our algorithm to identify tumors in the CE images and achieved a good accuracy. Moreover, to the best of our knowledge, the powerful Multiple instance learning method or any of its variants have never been used in the field of medical imaging even though it holds a lot of promises.