3D Mosaicing from Stereo Video
I will develop an algorithm to register range images taken from a stereo pair using image based techniques. Stereo data of a scene from different viewpoints is obtained by moving around a caliberated camera pair. From the stereo & image data, the system solves for the rigid body motion between the various camera pair locations. This information can be used to register all the range images wrt to a common reference frame. This gives us a planar mosaic of the scene with disparity values at each pixel, which can be used to render views from new viewpoints.
Szeliski [CVPR 99] uses a similar formulation using multiple views to obtain dense stereo depth maps. In his image mosaicing report [DEC 94], Szeliski presents a formulation to recover projective depth and register range images from an image sequence. In this work, he represents the depth map using a tensor product spline and recovers depth estimates only at the spline control vertices. Other related work in the rendering part includes the Layered depth images [SIGGRAPH 98] by Shade et al, View Interpolation [SIGGRAPH 93] by Chen & Williams and Plenoptic modeling [SIGGRAPH 95] by McMillan and Bishop. The stereo algorithm that I am currently using was developed by Zitnick and Kanade[CMU-RI-TR-99-35].
Details of the algorithm:
I model the transformation between two images taken from different viewpoints using 11 parameters (8 parameter for the 2D homography and 3 parameters for the translation along the epipole direction). Note that the projective depth at every pixel is known to us from the stereo data. I use the Levenberg-Marquardt algorithm to compute these unknown parameters by minimising the sum squared difference in the intensity values of the corresponding pixels.
Novelty of the algorithm:
A new technique to register range images using image based techniques.
A different approach to solve structure from motion?
Testing the algorithm:
I have taken a few stereo sequences of the NSH common area and I will try to mosaic the various images to produce a composite image mosaic with range information at each pixel.