Hybrid Approach for Efficient and Accurate Category-Agnostic Object Detection and Localization with Image Queries in Human-Robot Interaction

Abstract

Efficient and accurate object detection and localization play a crucial role in enabling robots to understand and interact with their environment. To this end, this paper presents a novel hybrid approach that combines deep learning and feature-based methods to address category-agnostic object detection and localization using image queries. By leveraging the strengths of both approaches, our method achieves superior performance in accurately localizing and segmenting objects, surpassing traditional feature-based template matching methods and the widely-used YoLov3. The proposed method utilizes a category-agnostic semantic segmentation framework, where objects are segmented based on their presence rather than their specific categories. Through quantitative evaluations on both synthetic and real-world datasets, our approach demonstrates remarkable accuracy and robustness in various scenarios, including objects with arbitrary shapes. The results demonstrate that the proposed approach provides an effective object detection and localization tool for visual servoing augmentation.