Contributions in Computational Video

As mentioned in the introduction, there are three steps in video stabilization and rolling shutter rectification. We are working primarily on the first two steps of motion estimation. We have developed an online algorithm for camera and gyroscope calibration without any prior knowledge of the devices. The camera motion can be reliably obtained from gyroscope after the calibration and synchronization is performed. For offline motion smoothing, we exploit the manifold structure of the sequence of 3D rotation matrices. Then we formulate the offline motion smoothing problem as a constrained manifold regression problem. The formulated problem is solved efficiently by an extension of the two-metric gradient projection method. For online motion smoothing, we develop a constrained multiple-model estimation algorithm to adaptively smooth the camera motion while guaranteeing that no black borders intrude the stabilized video frames.

The primary contributions of our research be summarized as follows.

  1. Online Camera-Gyroscope Auto-Calibration for Cellphones: In this contribution, we develop an online method that estimates all of the necessary parameters while a user is capturing video. This algorithm is based on an implicit extended Kalman filter (EKF). Each video frame provides a view of the 3D scene and triggers the update of the EKF through multiple-view geometry. By extending the recent proposed multiple-view coplanarity constraint of camera rotation to rolling shutter cameras, we propose a novel implicit measurement that involves only camera rotation but works for any camera translation, including zero translation (pure rotation). The implicit measurements can be effectively used in the EKF to update the estimate of state vectors. This algorithm is able to estimate the needed calibration and synchronization parameters online with all kinds of camera motion, and can be embedded in video stabilization for fast camera motion estimation using gyroscopes. Both Monte Carlo simulation and cellphone experiments show that this online calibration and synchronization method converges fast to the ground truth values.
  2. Constrained 3D Rotation Smoothing via Global Manifold Regression: In this contribution, we present a novel offline motion smoothing algorithm for video stabilization. We use a pure 3D rotation motion model with known camera projection parameters. We directly smooth the sequence of camera rotation matrices for the video frames by exploiting the Riemannian geometry on a manifold. We consider the entire set of sequences of rotation matrices as a Riemannian manifold. This allows formulation of the offline motion smoothing problem globally as a regression problem on the manifold based on geodesic distance. We introduce a geodesic-convex constraint on the manifold to approximate black-border constraint so that the entire motion smoothing problem is kept geodesic-convex on the manifold. To solve the formulated constrained smoothing problem on the manifold, we compute the gradient and Hessian of the objective function using Riemannian geometry, and then extend the two-metric projection algorithm in Euclidean space to non-linear manifolds. The geodesic-distance-based smoothness metric better exploits the manifold structure of sequences of rotation matrices. The geodesic-convex constraints effectively guarantee that no black borders intrude into the stabilized frames. The proposed manifold optimization algorithm can find the global optimal solution in only a few iterations. Experimental results show that the proposed motion smoothing method outperforms state-of-the-art methods by generating more stable videos with less distortion.
  3. Real-time Motion Smoothing via Constrained Multiple-Model Estimation: In this contribution, we develop a real-time motion smoothing method. This method is motivated by Kalman filtering-based motion smoothing with a constant velocity model. We use estimate projection to ensure that the smoothed motion satisfies black-border constraints, which are modeled exactly as linear inequalities for general 2D motion models. Then we combine the estimate projection with Bayesian multiple-model estimation to achieve adaptive smoothing in a probabilistic way. Experimental results show how the proposed algorithm can better smooth the camera motion and stabilize videos in real time.
Note: The above description came from the 2014 PhD Dissertation by Dr. Chao Jia at The University of Texas at Austin.

Contributions in Computational Photography

For digital still cameras, we have automated selected rules of photographic composition to help amateur photographers take better pictures. The automated methods rely on an auto-focus filter, software-controlled shutter aperture, and image processing algorithms to detect and segment the main subject of the photograph. The image processing algorithms are low-complexity and amenable to fixed-point implementation:
  1. Main subject detection and segmentation
  2. Photographic composition rules using main subject segmentation
In conducting this research, we have kept in mind from the beginning that the algorithms we develop will ultimately be implemented on a fixed-point digital signal processor. We are currently mapping the above algorithms onto fixed-point digital signal processors.


Mail comments about this page to bevans@ece.utexas.edu.