Reconstructing 3D dynamic scenes with non-accidentalness principle

Relative scale ambiguity is one of the most important problems we encounter while reconstructing dynamic scenes from monocular image sequences. As an example, consider a scene where a car is moving on the road and a moving camera is recording it. By just looking at the images, we can not decide whether it is a real car on the road or a small toy car hovering in front of the camera. In this paper, we first analyze this problem and then present our solutions which are based on the non-accidentalness principle.