Enhancement of Unsupervised Object Detection using Supervised Method

Abstract

Reasoning in terms of objects is a critical skill for any intelligent agent. To reason in terms of objects, an agent needs a method for discovering and detecting objects in the visual world. This is achieved through unsupervised object detection. Unsupervised object detection involves the identification of a large number of objects within images, without any supervision or labels. This area has received significantly less attention as compared to its supervised counterpart, which has achieved human-level accuracy. In a lot of cases, labeling each object in the image is impractical and requires a lot of labor. While unsupervised techniques are suitable for the task of object localization, the learner cannot identify the classes of the localized objects unless it is explicitly specified. Combining supervised learning techniques with unsupervised learning techniques can improve the performance of existing algorithms for both detection and classification. In this paper, we have studied a state-of-the-art unsupervised object detection approach and extended it for object classification, by combining it with supervised techniques.