Research On Sparse-Feature-Based Rendering

With the vigorous development of3D technology,3DTV becomes more and more popular in everyone’s life. In comparison against wearing glasses3DTV, naked eyes3DTV helps the audience get rid of the glasses and thus provides a better visual experience. It’s a common way to use view rendering technology to synthesize the virtual views which are needed in the naked eyes3DTV. A novel sparse-feature-based rendering algorithm uses sparse but reliable feature points as disparity guidance and solves mapping relationship between virtual views and original views through the global optimization and thus avoids the technological difficulties which are depth generation and holes filling in the traditional method called DIBR.However, the current solution mostly focuses on the images not videos, and is far from real-time. This paper is concentrated on this sparse-feature-based rendering algorithm and the main achievements include:1. A whole framework of sparse-feature-based rendering algorithm is proposed and the sub modules are studied and optimized to get a better real-time performance, such as temporal continuity estimate, visual saliency, and the construction and solving of energy equation.2. A two-layer descriptor and two-step matching for stereoscopic images is proposed which improves the structure of the SURF descriptor and uses the range of the disparity to guide the matching process and thus reduces the time of description generation and matching. And a kind of feature extraction algorithm for stereo video is also proposed, which takes the advantage of fast operation time of the optical flow method, and ensure the accuracy and the number of feature points at the same time.3. An adaptive block of sparse-feature-based rendering algorithm is proposed which partitions the images according to their saliency. Different objects are partitioned into different blocks to avoid the influence of each other and blocks of the same objects are merged together to reduce the number of the block.