Research map / Video / Video segmentation and tracking
Video Frame Interpolation via Adaptive Separable Convolution
Simon Niklaus, Long Mai, Feng Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace L1-only training with L1 then perceptual (feature reconstruction) fine-tuning (ablation) | Ours - L1 · Middlebury interpolation (average) | interpolation error (all): 5.61 -> 5.81 | Hurt |
| replace L1-only training with L1 then perceptual (feature reconstruction) fine-tuning | Ours - L1 · Cross-validation (250,000 samples) | PSNR: 32.92 -> 32.37 | Hurt |
| replace L1-only training with L1 then perceptual (feature reconstruction) fine-tuning | Ours - L1 · See You Again video | PSNR: 41.31 -> 40.88 | Hurt |
| add perceptual loss fine-tuning to the separable convolution network (ablation) | Ours - L1 · visual quality (sharpness) | user study / qualitative | Improved |
| estimate separable 1D kernels instead of 2D kernels (AdaConv) | AdaConv · Middlebury interpolation (average) | interpolation error (all): 6.20 -> 5.61 | Improved |
| estimate separable 1D kernels instead of 2D kernels (AdaConv) | AdaConv · Cross-validation (250,000 samples) | PSNR: 30.16 -> 32.92 | Improved |
| estimate separable 1D kernels instead of 2D kernels (AdaConv) | AdaConv · 1080p interpolation speed | speed: over 20 times faster | Improved |
| use 1D separable kernels approximating 41x41 kernels | dense 2D kernel estimation · 1080p frame kernel memory | GB: 26 GB -> 1.27 GB | Improved |
| use four sub-networks for the four kernels instead of a joint representation (ablation) | kernel estimation network · training | convergence speed | Improved |
| use bilinear upsampling in decoder instead of poorly chosen upsampling (ablation) | decoder · interpolated frames | checkerboard artifacts | Improved |
| use L2 loss instead of L1 (ablation) | separable conv network · interpolated frames | visual sharpness | Hurt |
| synthesize frame directly instead of estimating kernels (ablation) | separable conv network · interpolated frames | visual quality | Hurt |
Closest papers by meaning
- Video Frame Interpolation via Adaptive Convolution (CVPR 2017 · similarity 0.96)
- Multiple Video Frame Interpolation via Enhanced Deformable Separable Convolution (TPAMI 2021 · similarity 0.93)
- Video Frame Interpolation via Deformable Separable Convolution (AAAI 2020 · similarity 0.92)
- Depth-Aware Video Frame Interpolation (CVPR 2019 · similarity 0.92)
- EDEN: Enhanced Diffusion for High-quality Large-motion Video Frame Interpolation (CVPR 2025 · similarity 0.88)
- Generative Inbetweening: Adapting Image-to-Video Models for Keyframe Interpolation (ICLR 2025 · similarity 0.88)
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.
