홈 › Papers

Multi-view Arbitrary Style Transfer

In this paper, we introduce pioneering algorithms for multi-view arbitrary style transfer. Multi-view arbitrary style transfer is an advanced study of the conventional monocular arbitrary style transfer, which aims to preserve the consistent style on the common region across the given arbitrary number of views. We intend to address the multi-view inconsistency problem by minimizing the difference in the color and feature values of the corresponding regions. However, the conventional feature extractors generally produce a feature vector of a point from its rectangular local patch, and such local patches are misaligned across the views in a multi-view environment due to various camera poses. Thus, even if we assimilate the feature vectors of the corresponding pixels, since the spatial distribution of the surrounding feature vectors within their local patches are different and decoding such misaligned patches induces misaligned brushstrokes and geometric patterns, these feature vectors can be decoded to distinctive style texture. Based on the observation, we intend to interpret this challenging problem in terms of the photometric inconsistency and the stroke inconsistency. We propose a photometric consistency loss, which directly enforces the geometrically consistent style texture across the view, and a stroke consistency loss, which matches the characteristics and directions of the brushstrokes by aligning the local patches of the corresponding pixels before minimizing feature deviation. Then, We construct an optimization-based multi-view arbitrary style transfer framework (MVAST-O) with photometric and stroke consistency losses and extend it to a feed-forward framework (MVAST-FF) to overcome the chronic computational inefficiency issue of optimization-based algorithms. We validate our methods on the DTU dataset, a large-scale multi-view stereo dataset, and confirmed the superiority on preserving appearance consistency throughout the stylized multi-view images.

Code (0)

등록된 구현이 없습니다.

Tasks

Similar Papers 제목 키워드 기반