I'm trying to reconstruct objects from Internet photographs. In most cases, I have only two images, so I'm looking for methods that can compute 3D points from a pair of images. I have the following code, based on this tutorial, which seems to compute a reasonable¹ fundamental matrix: def readMatches(...): # Returns a set of (manually annotated) point matches # sufficient for computing the fundamental matrix. def computePose(pts1, pts2): pts1 = np.int32(pts1) pts2 = np.int32(pts2) F, mask = cv.findFundamentalMat(pts1, pts2, cv.FM_LMEDS) pts1 = pts1[mask.ravel() == 1] pts2 = pts2[mask.ravel() == 1] return pts1, pts2, F if __name__ == '__main__': pts1, pts2 = readMatches(...) pts1, pts2, F = computePose(pts1, pts2) (¹ At least, drawing the epipolar lines using the code in the tutorial produces a result that looks reasonable.) What is the next step to go from this to a 3D point cloud of the input points? (If it helps, I can provide a set of point pairs for one of the inputs that should form a series of parallel lines. That is all I have to go on as far as camera intrinsics, however.) Python OpenCV code is preferred. Please refrain from posting answers that require purchasing a book. Continue reading...
