Openset Approach To Object Detection, An

Abstract

Object detection has been a popular area of research in computer vision, where during the recent years deep learning based approaches have taken over the traditional approaches based on hand engineered features like HOG and SIFT. A common aspect in all of these recent approaches has been the approach taken to identify parts of an image as backgrounds. This approach generally limits the performance of such a deep learning approach when it is applied outside the academic datasets into the real world. We highlight this performance gap by simulating a real-world detection problem using our novel openset object detection protocol. In this work, we present a new approach to handle the backgrounds while reintroducing object detection as an openset problem. In the recent past some openset approaches have been proposed, but their applications have been limited to image classification problems. In this thesis, we apply one such existing openset approach to the problem of object detection while proposing our novel openset approach of MultiModel OpenMax. We find that our proposed approach works better than the earlier openset approaches but is still insufficient to solve the openset aspect of object detection.