Fast Algorithm for Disparity Estimation in ATSC-M/H based Hybrid 3DTV

Abstract

ATSC-M/H based hybrid 3DTV, which is one of the service compatible 3DTV system, has considerable quality gap between the left and right views. And CRA(Conditional Replenishment Algorithm) has been proposed to deal with the issue of resolution mismatch and improve the visual quality. In CRA, the disparity vectors of stereoscopic images are estimated. The disparity compensated left view and simply enlarged right view are compared and conditionally selected for generating the enhanced right view. In order to implement CRA, a fast algorithm is strongly required because the disparity vectors need to be obtained at every layer and the complexity of CRA is quite high. In this paper, we adopted SDSP(Small Diamond Search Pattern) instead of full search and predicted the initial position of search pattern by examining the spatio-temporal correlation of disparity vectors and also suggested the SKIP mode to limit the number of processing units. The computer simulation showed that the proposed fast algorithm could greatly reduce the processing time while minimizing the quality degradation of reconstructed right view.

A systematic analysis of regression models for protein engineering

“…In an ablation study we considered whether an unsupervised dimensionality-reduction pre-processing of the averaged representations could have a positive effect on regression performance, but generally found this to be detrimental (see Fig W in S1 File ). However, we note that several recent protein optimization studies have found it beneficial to include in their optimization protocol a supervised dimensionality reduction step trained specifically on the task [ 30 , 31 ].…”

“…with a density model p ( X )), or 3) employing a regression model which will associate out-of-domain predictions with high degrees of uncertainty. Since the topic of our paper is regression, we will focus on the last point here, but note that examples of the first two options exist in recent protein optimization methods, where lists of candidates are generated to be close to wildtype proteins, for instance using a generative model of p ( X ) [ 30 , 31 ].…”

“…However, in computational labs such experiments might not be accessible, which has led research towards finding computational proxies. One approach that has been suggested is to train two oracles on subsets of the data, optimize against one, and validate against the other—averaging over both permutations [ 31 ]. Ideally, the model assumptions should be different between the two oracles, to ensure that extrapolation artefacts are different between them.…”