Depth Estimation for 3D Object Inspection Using RGB Videos

10.29428/9789860544169.201801.0088

Jing-Min Chen；Hsia-Yuan Lin；Shyi-Chyi Cheng

depth estimation ； pose detection ； 3D object reconstruction ； deep learning

NCS 2017 全國計算機會議

2017（2018 / 01 / 01）

459 - 464

英文

The usage of a drone for capturing RGB videos has become a popular manner to inspect the completeness of a 3D object, in which the key task is to produce 3D contents by estimating the depth for the frames. To achieve the goal, many 2D to 3D video conversion algorithms have been proposed. However, using a RGB video, conventional depth estimation algorithms are not applicable to inspect the condition of an object in 3D domain since these algorithms are limited by the requirement of an initial accurate depth map and the propagation error of depth information across frames. To address this problem, we propose a two-stage model-based depth estimation method for automatic 2D-to-3D conversion. In the first stage, the pose class of the object under surveillance is detected using a deep-learning architecture. In the second stage, the depth image of the test frame is generated by transforming the pre-stored depth map of the representative object according to the correspondences between the test object and the representative object. The accurate depth information is then used to generate the 3D contents for object inspection. To verify the effectiveness of the proposed method, a 3D object reconstruction system is constructed. Finally, we verify the ability of the established reconstruction system on publicly available benchmark datasets, and compare it with the sate-of-the-art pose estimation algorithms. Experimental results show that our approach outperforms the compared methods on the accuracy of pose estimation.