earticle

영어

We address data collection issues in monocular 3D object detection. Since deep learning based object detection systems often require a variety of human annotations for advanced functionalities and also a lot of training datasets for better performance, it is increasingly important for researchers to review currently available databases, and prepare their own datasets with customized labels for target applications such as monocular 3D object detection. On top of the survey of related datasets, we study two different methods of collecting new datasets. As an example of our methods, we present five realworld object instances for pure RGB-based cuboid detection, and simulate random scenes with the objects for training and testing. In order to improve detection accuracy, all the simulated objects with limitless numbers and varied sizes can appear in high resolution images. As a baseline model, we validate our datasets using the YOLO-like standard deep learning architecture. In a coarse-to-fine manner, annotations such as cuboid, segmentation masks, and 3D models are first accessible on a down-sampled version of the simulated image and then on a sequence of higher resolution sub-images possibly utilized. Finally, we discuss experimental results and future directions.