Film and consumer electronics industries have known in the last few years huge technological improvements to capture, transmit and display high-quality monoscopic and stereoscopic video content. These improvements aim at providing to the viewer the most realistic viewing experience. Due to artistic intentions or physical limitations to efficiently capture and transmit video contents, it is sometimes necessary to combine simultaneously captured and synthetic data while taking care to maintain a photo-realistic rendering. To efficiently process captured and synthetic content simultaneously, production and post-production operators need to be assisted by sophisticated automatic tools. Among these tools, we thoroughly investigated both view synthesis quality assessment and long-term dense motion estimation issues. 3D autostereoscopic displays rely on the generation of realistic-looking virtual viewpoints through disparity estimation and view interpolation involved together within Depth-Image-Based Rendering (DIBR) algorithms. Despite recent advances, DIBR algorithms do not always provide artifact-free synthesized views and induce new types of artifacts whose impact can be harmful for the observer. Our contribution in this context has been to develop and evaluate a new full-reference objective image quality assessment metric dedicated to view synthesis quality assessment. Also required by recent applications such as scene segmentation or dynamic scene analysis techniques, long-term dense displacement fields allow to propagate synthetic data to the whole sequence in the context of high quality video editing. However, state-of-the-art optical flow estimators show strong limitations toward long-term requirements since classical optical flow assumptions are not valid for non-consecutive frames. Therefore, we proposed several contributions to long-term dense motion estimation based on multi-step optical flow vectors. First, a sequential fusion approach including a spatio-temporal multilateral filtering has been investigated toward long-term dense correspondences robust to temporary occlusions. Then, an alternative method has been studied based on combinatorial integration and statistical selection. Finally, we proposed multi-reference frames strategies to correlate trajectories estimated with respect to multiple reference frames selected according to motion quality criteria. Our contributions in both contexts offers new perspectives, especially for joint stereo and motion processing. In this direction, an automatic disparity correction framework using long-term dense displacement fields has been addressed.
