Depth-Based Region Proposal: Multi-Stage Real-Time Object Detection

Abstract

Many real-time object recognition systems operate on two-dimensional images, degrading the influence of the involved objects' third-dimensional (i.e., depth) information. The depth information of a captured scene provides a thorough understanding of an object in fulldimensional space. During the last decade, several region proposal techniques have been integrated into object detection. scenes’ objects are then localized and classified but only in a two-dimensional space. Such techniques exist under the umbrella of two-dimensional object detection models such as YOLO and SSD. However, these techniques have the issue of being uncertain that an object's boundaries are properly specified in the scene. This paper proposes a unique region proposal and object detection strategy based on retrieving depth information for localization and segmentation of the scenes’ objects in a real-time manner. The obtained results on different datasets show superior accuracy in comparison to the commonly implemented techniques with regards to not only detection but also a pixel-by-pixel accurate localization of objects.