Customized Obstacle Dataset Generation from RGB-D Datasets

Abstract

The field of Obstacle Detection in computer vision is evolving rapidly, with many researchers contributing innovative solutions to address different challenges. However, the current landscape presents several barriers that hinder effective comparison and collaboration among researchers. A significant challenge arises from the limited availability of publicly accessible obstacle datasets, with existing datasets predominantly focused on specific contexts such as obstacles on roads, sidewalks, or particular industrial settings. The varied application contexts, environments, and hardware equipment in use compel researchers to develop and use custom datasets tailored to their application-specific requirements. Constraints such as time, funding, and limited access to resources often discourage researchers from re-creating new datasets for the purpose of comparative evaluations, resulting in incomplete insights into the strengths and weaknesses of different approaches. To address this deficiency in comparative analysis, we propose a novel method for generating obstacle datasets from existing RGB-D datasets dynamically. The proposed method offers a quick and effortless way to generate customized variants of Obstacle dataset, without having to manually build them from the ground up. The cross-dataset compatibility of the proposed method further enhances the versatility and wide adoption potential of the proposed method. © 2024 IEEE.