Key points
- AnimeInterp introduces segment-guided matching to handle anime's flat colour where texture-based flow fails.
- A recurrent flow-refinement module addresses the large non-linear motions common in anime.
- It released ATD-12K, the standard 12,000-triplet benchmark for 2D animation inbetweening.
Summary
Li et al. (2021) addressed the limitations of general video interpolation methods on 2D animation, where flat colours, sharp outlines and large non-linear motions cause texture-based optical flow to fail. AnimeInterp introduces a segment-guided matching module that exploits anime's region structure and a recurrent flow-refinement module for large-displacement handling, together producing accurate inbetween frames. The paper also released ATD-12K, a benchmark of 12,000 anime triplets that became the standard evaluation dataset for 2D animation interpolation. AnimeInterp is the founding domain paper for deep anime inbetweening and provides the methodological and benchmark baseline for subsequent methods including the diffusion-based ToonCrafter.
Related items
Source
Source: CVPR 2021 (IEEE/CVF) ↗ (Research)
Cite this item
Li, S., Zhao, S., Yu, W., Sun, W., Metaxas, D. N., Loy, C. C. & Liu, Z. (2021). ‘Deep Animation Video Interpolation in the Wild (AnimeInterp)’, CVPR 2021 (IEEE/CVF). Available at: https://arxiv.org/abs/2104.02495
Your reference manager can also read this page directly: with the Zotero (or Mendeley) browser connector installed, save it straight to your library. Whole-collection exports: RIS, BibTeX, CSL-JSON.