Unsupervised Detection of Unclassified Objects via Graph-based Segmentations and Image Saliency Fusion for Automated Incremental Learning

Abstract

Current state-of-the-art machine learning systems efficiently identify known classes of objects with a high rate of success. However, their inference capabilities become limited in the presence of objects for which little or no training data exists, or when extreme environmental conditions drastically affect the scene appearance. Therefore, their effectiveness when deployed in the context of exploring and interacting with previously unencountered environments with novel entities, is often limited. This research is focused on complementing such operations by discovering previously unclassified objects, while also refining their localization in images to facilitate robust tracking and extraction. The ultimate purpose is to facilitate the automated enhancement of the trained classification system with new class information via incremental retraining. The proposed unclassified object detection pipeline involves a multi-step architecture that starts by generating object proposals from a hierarchical set of graph segmentations across multiple color and frequency domains. Subsequently these are filtered, further refined, and fused by exploiting visual attention prediction models that estimate pixel-wise saliency in color images. This bottom-up approach aims to remain generic, while managing to propose a meaningfully small set of high quality regions within the image context so as to optimize their evaluation for retraining. The presented unclassified object discovery results come from experimental datasets taken from the viewpoint of an aerial robot in a variety of operating conditions and environments. They are given with reference to the YOLOv3 state-of-the-art detection algorithm to indicate the complementarity of our approach in managing to isolate unknown entities during real-time operation.