3D models from extended uncalibrated video sequences: addressing key-frame selection and projective drift

Abstract

In this paper, we present an approach that is able to reconstruct 3D models from extended video sequences captured with an uncalibrated hand-held camera. We focus on two specific issues: (1) key-frame selection; and (2) projective drift. Given a long video sequence it is often not practical to work with all video frames. In addition, to allow for effective outlier rejection and motion estimation it is necessary to have a sufficient baseline between frames. For this purpose, we propose a key-frame selection procedure based on a robust model selection criterion. Our approach guarantees that the camera motion can be estimated reliably by analyzing the feature correspondences between three consecutive views. Another problem for long uncalibrated video sequences is projective drift. Error accumulation leads to a non-projective distortion of the model. This causes the projective basis at the beginning and the end of the sequence to become inconsistent and leads to the failure of self-calibration. We propose a self-calibration approach that is insensitive to this global projective drift. After self-calibration triplets of key-frames are aligned using absolute orientation and hierarchically merged into a complete metric reconstruction. Next, we compute a detailed 3D surface model using stereo matching. The 3D model is textured using some of the frames.