DragView is an interactive framework that generates novel views of unseen scenes from a single source image, using a sparse set of unposed multi-view images. The user drags a source view through a local coordinate system, and pixel-aligned features are obtained by projecting sampled 3D points onto the source view. The framework incorporates a view-dependent modulation layer to handle occlusion during projection. It broadens the epipolar attention mechanism to include all source pixels, facilitating the aggregation of point features from other unposed views. A transformer is used to decode ray features into final pixel intensities. DragView demonstrates superior performance in view synthesis quality and is user-friendly.
Publication date: 5 Oct 2023
Project Page: https://zhiwenfan.github.io/DragView/
Paper: https://arxiv.org/pdf/2310.03704
