IIIT Hyderabad Publications |
|||||||||
|
6DOF POSE AND SHAPE ESTIMATION OF OBJECTS FROM MONOCULAR IMAGESAuthor: ANIKET DINESH POKALE Date: 2021-04-30 Report no: IIIT/TH/2021/48 Advisor:Madhava Krishna AbstractAs more robots continue to pervade our workplaces, homes, and lives, developing accurate and robust algorithms for their effective and safe operation is of paramount importance. We see more and more of such robots being used for various applications and we certainly have a new future ahead of us filled with smart robots navigating autonomously, be it a vacuum cleaner robot or logistics drone or an autonomous car. There is a plethora of ongoing research work trying to improve upon various algorithms for autonomous navigation of robots. Object detection is one such research area which has gained significant momentum. The 6 degree of freedom object pose estimation enables the robot to localize the object in the surroundings and perform its task accordingly. We see that there have been some very interesting work done for 6DOF pose estimation from monocular images. In this thesis we propose a new method for 6DOF object pose estimation using a single monocular camera with the assumption of known camera height. We detect the objects in the RGB images using deep learning and use a renderer called neural renderer to optimize for the 6DOF pose of the object. Furthermore we also explore how object detection can aid monocular SLAM. Here we propose the first most principled formulation of it’s kind which incorporates the object optimization into the bundle adjustment of SLAM, we call it the joint bundle adjustment. This joint formulation allows us to optimize the object shape, pose, and also the 3D semantic points and camera localization in a joint fashion. Full thesis: pdf Centre for Robotics |
||||||||
Copyright © 2009 - IIIT Hyderabad. All Rights Reserved. |