Article ID Journal Published Year Pages File Type
4948189 Neurocomputing 2016 21 Pages PDF
Abstract
In recent years, the understanding the semantics of 3D scenes has been a wide interesting researching point in many application. However, 3D scenes detection remains many problems, due to the difficulty in acquiring sufficient 3D model towards training effective classifiers. In order to address these problems, in this paper, we first publish a new real-world 3D model dataset MV-RED, which includes 505 objects recoded by Kinect camera. Then we propose a novel 3D object detection approach in real-world scenes combined RGB image based on MV-RED dataset. In order to improve the detection precision, we also utilize the tracking method to improve the detection results. Finally, we evaluate our approach on the RGB-D dataset which is provided by Lai et al. (2012) [20], achieving much greater efficiency and comparable accuracy. Our approach shows further major gains in accuracy when the training data from the target scenes is used, outperforming state-of-the-art approaches with far better efficiency.
Keywords
Related Topics
Physical Sciences and Engineering Computer Science Artificial Intelligence
Authors
, , , ,