Recovery of Relative Aane Structure Using the Motion Flow Field of a Rigid Planar Patch

Recently there has been much interest in 3D reconstruction from uncalibrated views. Relative AAne Structure has been proposed 1] as an alternative framework to projective reconstruction. Proposed algorithms in 1, 2] were based on an oo-line feature corresponding process. This paper extends the relative aane structure from motion to a `correspondenceless' 1 framework. A `projected planar patch ow eld (ppow)''3], deened by the image motion of a rigid planar patch, is used to estimate the eight pure parameters 4] that describe the 2D projective transformation between the current and reference frame. Thèppow registration' 3] brings into very precise alignment any two pictures of the same plane, even if taken by diierent cameras with diierent calibration. However, registration will fail for objects in the scene that are not on the particular chosen plane. By exploiting this registration failure, we may determine the relative aane structure of the scene. Applying`ppow registration' to an image sequence, one will observe that in the registered sequence objects on the plane of interest are not moving, while those that are not on the plane will move with a velocity proportional to their distance from the plane and inversely proportional to their depth in the camera coordinates. The optical ow between the registered images is an epipolar eld and yields to the recovery of epipolar geometry and relative aane structure. Results on a sequence of real images helps the understanding of the process and shows that this approach is quite promising.