Marker-less 3D Pose Estimation

Jeffrey Resodikromo

Department of Information and Computing Sciences, Utrecht University


This thesis presents the research done on a marker-less pose estimation method. The method by Gall et al. has been implemented, analysed and improved for use by Noldus IT. The method consists of a combination of two optimization approaches, namely a local and a global optimization. For the local optimization a weighted least squares problem is solved to minimize the distance between the model and the silhouettes. The global optimization uses a combination of particle filter and simulated annealing, called the Interacting Simulated Annealing, to estimate the pose where the local optimization fails. Given the mesh model, camera calibration data and the camera video sequences an accurate pose is estimated. A full analysis is performed on the method and based on the evaluation results, improvements were made to increase the quality and decrease the runtime of the estimation. A successful dynamic threshold was added to create a seamless connection between the two optimization approaches. Adding skeleton correspondences to the local optimization that takes the inner structure of the model into account, increases the quality of our estimation. Our implementation provides an efficient and accurate pose estimation for use in practical applications in a controlled environment and gives the basis for extensions.


Marker-less 3D Pose Recognition [PDF]
Jeffrey Resodikromo

System Pipeline

Pose estimation:

Without the texture correspondences.

Including the texture correspondences.

Including the dynamic threshold and the skeleton correspondences. Without the texture correspondences. Increased runtime by factor 6.