The target of this research is to be able to automatically extract the 3D shape of a scene filmed by a common commercial camera freely held by a user. The velocity of the scene points projected on the camera image plane (optical flow) is an important source of information about the motion on the camera itself and the distance of the scene points. We use the so called differential epipolar constraint to compute first the velocity (both linear and angular) of the camera and then to extract the 3D positions of the scene points. The new algorithm that completely automatically performs these operations. The results are extremely good both for robustness and precision of the reconstruction.
A Comparison of Stereo Based and Flow Based Structure from Parallax M. Zucchelli, H. Christensen SIRS 2000
Flow Based Structure from Parallax Using Short Sequences with Automatic Rescaling M. Zucchelli, H. Christensen Submitted to ICCV2001
The Interplay between Structure Distortion and Motion Bias induced by Calibration Errors M. Zucchelli, J.Kosecka Submitted to ICCV2001
An Application of Machine Learning to the Optimization of Disparity Maps G. Cervone, M. Zucchelli Submitted to IJCAI01
The images show the reconstruction of a teabox performed with 10 frames acquired at 30Hz by a freely moving camera. The total sequence length is 0.3 seconds and the total disparity about 10 pixels.
|
|
|