Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Voting-Based Pose Estimation
See description
The classic method for camera pose estimation is to embed a sampling strategy into a RANSAC framework and use Minimal Solvers for computing a pose hypothesis that are later verified in a separate step. The complexity of this methodology is determined by the inlier likelihood and number of points that are needed for a minimal sample, which can be a problem. In contrast, voting based approaches are of linear complexity in the number of correspondences and not affected by inlier probabilities.
The goal of this work is to implement a basic methodology following [1]. Here we must rely on a known gravity direction to guarantee a 2-dimensional voting space. Apart from the beneficial time complexity voting based methods naturally deliver multiple solutions.
Furthermore -- and maybe even more importantly -- the general approach extends naturally to a deep learning framework. Again, this contrasts other pose estimation algorithms for which the training phase becomes artificial and complex, eg. [2].
[1] Zeisl et al. "Camera Pose Voting for Large Scale Image-Based Localization", ICCV 2015
[2] Kendall et al. “PoseNet: A Convolutional Network for Real-Time 6-DOF Camera Relocalization”, ICCV 2015
The classic method for camera pose estimation is to embed a sampling strategy into a RANSAC framework and use Minimal Solvers for computing a pose hypothesis that are later verified in a separate step. The complexity of this methodology is determined by the inlier likelihood and number of points that are needed for a minimal sample, which can be a problem. In contrast, voting based approaches are of linear complexity in the number of correspondences and not affected by inlier probabilities. The goal of this work is to implement a basic methodology following [1]. Here we must rely on a known gravity direction to guarantee a 2-dimensional voting space. Apart from the beneficial time complexity voting based methods naturally deliver multiple solutions. Furthermore -- and maybe even more importantly -- the general approach extends naturally to a deep learning framework. Again, this contrasts other pose estimation algorithms for which the training phase becomes artificial and complex, eg. [2]. [1] Zeisl et al. "Camera Pose Voting for Large Scale Image-Based Localization", ICCV 2015 [2] Kendall et al. “PoseNet: A Convolutional Network for Real-Time 6-DOF Camera Relocalization”, ICCV 2015
Explore a voting based method for pose estimation and compare the performance to traditional RANSAC based approaches.
Explore a voting based method for pose estimation and compare the performance to traditional RANSAC based approaches.