Efficient manifold preserving edit propagation with adaptive neighborhood size

Recent manifold preserving edit propagation (Chen et al., 2012) [1] provides a robust way for propagating sparse user edits to a whole image or video, which preserves the manifold structure formed by all pixels in feature space during edit propagation. However, it consumes a big amount of time and memory especially for large images/videos, limiting its practical usage. In this paper, we propose an efficient manifold preserving edit propagation method. We accelerate the original method from two aspects. First, instead of using a fixed neighborhood size in building the manifold structure, we adaptively determine neighborhood size for each pixel based on its local complexity in feature space, which largely reduces average neighborhood size. Secondly, following Xu et al. (2009) [2], we adaptively cluster all pixels, and solve the edit propagation problem on clusters instead of pixels. Our experiment shows that, compared to the original method (Chen et al., 2012) [1], our method significantly reduce time and memory costs without reducing visual fidelity.