Atsushi KOIKE

The authors study an objective method to estimate perceptual picture quality of the coded pictures based on No Reference framework. The proposed method extracts two image features that reflect the significance of blocking and flickering artifacts and estimates the subjective quality by weighted sum of the image features. Computer simulations show that the proposed method can estimate subjective quality at a correlation coefficient of 0.754.

In this research, we address the problem of 3-D model generation from disparity maps. Given an inaccurate 3-D model, dense multiview images are generated. Disparity maps between each pair are estimated using stereo matching algorithm. The disparity maps are projected into space as depth candidates. The kernel classifier is applied to the depth candidates for each layer of candidates from ground, and the candidates with higher probability values are selected for further fine-tuning process. In fine-tuning step, the best location of the border of the objects in each layer from ground is determined using dynamic programming. Finally the 3-D model is generated. Using the generated 3-D model, we regenerate the dense multiview images and disparity maps. We perform the same process for several iterations until the changes in the generated 3-D model are small. Experimental result shows we can eventually enhance the quality of the 3-D model within a few iterations in comparison with the starting 3-D model.

We proposed divided ray space based method generating free viewpoint video using circular arrayed multi-viewpoint video. However, there were vertical gaps in free viewpoint videos, since vertical disparity cannot be acquired. So, by using new format of ray space described by horizontal disparity values that we have proposed in order to compress ray space data, vertical disparity can be realized exactly since vertical disparity can be calculated with the horizontal disparity.