Research map / Video / Video segmentation and tracking
Thin-Plate Spline-based Interpolation for Animation Line Inbetweening
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a raster-image animation line inbetweening pipeline that uses thin-plate spline coarse motion from keypoint correspondences, a flow refinement module, and a UNet, plus a new Weighted Chamfer Distance metric.
Problem
Animation line inbetweening: predicting intermediate line-art frames between key frames, where sparse pixels and large motion cause line disconnections and existing Chamfer Distance evaluation is misleading.
Approach
Model coarse large motion with a thin-plate spline warp driven by matched keypoints (from a frozen GlueStick matcher), refine it with an IFBlock-style flow module and a UNet, and evaluate with a perceptually aligned Weighted Chamfer Distance (WCD) plus EMD and a user study.
Key results
- On MixiamoLine240 test set with frame gap 9, the method achieves WCD 9.48 and EMD 3.37, versus AnimeInbet's 11.54 WCD, while AnimeInbet has the lower CD (17.76 vs 18.09).
- User study (20 participants) preferred the proposed method in 70.50% of comparisons overall, versus 19.40% for AnimeInbet and 10.10% for EMA-VFI.
- Ablation on the validation set: removing the TPS module (w/o MRM) raises CD from 9.235 to 11.800 and EMD from 1.343 to 4.233.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the motion refinement module to TPS-only coarse motion estimation (ablation) | Ours · MixiamoLine240 validation, N=5 | CD (×10^5): 11.800 -> 9.235 | Helped |
| add the motion refinement module to TPS-only coarse motion estimation (ablation) | Ours · MixiamoLine240 validation, N=5 | WCD (×10^4): 8.850 -> 7.994 | Helped |
| add the motion refinement module to TPS-only coarse motion estimation (ablation) | Ours · MixiamoLine240 validation, N=5 | EMD (×10^3): 4.233 -> 1.343 | Helped |
| add flow refinement (ablation) | Ours · MixiamoLine240 validation, N=5 | CD (×10^5): 9.314 -> 9.235 | Helped |
| add flow refinement (ablation) | Ours · MixiamoLine240 validation, N=5 | WCD (×10^4): 8.015 -> 7.994 | Helped |
| add flow refinement (ablation) | Ours · MixiamoLine240 validation, N=5 | EMD (×10^3): 2.625 -> 1.343 | Helped |
| use Ours instead of the strongest baseline | Animation line inbetweening · MixiamoLine240 test, N=1 | WCD (×10^4): 7.31 -> 7.28 | Helped |
| use Ours instead of the strongest baseline | Animation line inbetweening · MixiamoLine240 test, N=5 | WCD (×10^4): 9.05 -> 8.39 | Helped |
| use Ours instead of the strongest baseline | Animation line inbetweening · MixiamoLine240 test, N=9 | WCD (×10^4): 10.09 -> 9.48 | Helped |
| use Ours instead of the strongest baseline | Animation line inbetweening · MixiamoLine240 test, N=5 | CD (×10^5): 10.74 -> 10.79 | Hurt |
| use Ours instead of the strongest baseline | Animation line inbetweening · MixiamoLine240 test, N=9 | CD (×10^5): 17.76 -> 18.09 | Hurt |
Surprising findings
- CD can be misleading for line art: erasing random lines from a predicted circle lowers CD, so CD rewards disconnections.
- AnimeInbet achieves the best CD in several settings even though it shows noticeable missing lines, while the proposed method has the best WCD.
Limitations
- Does not achieve the best CD score in several settings, where AnimeInbet is better.
- Trained and evaluated only on the MixiamoLine240 dataset as reported.
Experimental setup
Datasets: MixiamoLine240
Benchmarks: MixiamoLine240
Metrics: CD WCD EMD user preference PSNR SSIM
Baselines: RIFE EMA-VFI AMT PerVFI EISAI AnimeInbet
Builds on: IFBlock (RIFE) Chamfer Distance GlueStick LPIPS
Closest papers by meaning
- Framer: Interactive Frame Interpolation (ICLR 2025 · similarity 0.84)
- SplineGS: Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video (CVPR 2025 · similarity 0.84)
- Arbitrary Motion Style Transfer with Multi-condition Motion Latent Diffusion Model (CVPR 2024 · similarity 0.84)
- FlexiClip: Locality-Preserving Free-Form Character Animation (ICML 2025 · similarity 0.84)
- Generative Inbetweening through Frame-wise Conditions-Driven Video Generation (CVPR 2025 · similarity 0.84)
- Align Your Gaussians: Text-to-4D with Dynamic 3D Gaussians and Composed Diffusion Models (CVPR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.