Research map / Architectures and efficiency / CNN design
SKFlow: Learning Optical Flow with Super Kernels
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| use super kernel (15x15) decoder instead of normal 3x3 kernel | GMA · Sintel train Final (C+T) | EPE: 2.74 -> 2.46 | Helped |
| increase large kernel size from 7 to 15 (ablation) | SKFlow · Sintel train Final / KITTI | EPE / Fl-all: Sintel final 2.70 -> 2.55; KITTI Fl-all 16.11 -> 16.26 | Mixed |
| increase large kernel size from 15 to 31 (ablation) | SKFlow · Sintel train Final (C+T) | EPE: 2.55 -> 2.65 | Hurt |
| add residual connections (ablation) | SKFlow super kernel block · KITTI-15 train (C+T) | Fl-all: 17.33 -> 15.47 | Helped |
| add auxiliary small kernel (ablation) | SKFlow super kernel block · KITTI-15 train (C+T) | Fl-all: 16.55 -> 15.47 | Helped |
| conical design instead of parallel (ablation) | SKFlow super kernel block · KITTI-15 train (C+T) | Fl-all: 17.82 -> 15.47 | Helped |
| conical design instead of funnel (ablation) | SKFlow super kernel block · KITTI-15 train (C+T) | Fl-all: 16.97 -> 15.47 | Helped |
| auxiliary kernel 1x1 instead of 3x3 (ablation) | SKFlow · KITTI-15 train (C+T) | Fl-all: 15.56 -> 15.47 | Mixed |
| apply super kernels in both updater and motion encoder instead of one (ablation) | SKFlow · Sintel train Final (C+T) | EPE: 2.65 / 2.56 -> 2.46 | Helped |
| deeper layers instead of super kernels (ablation) | GMA decoder · FlyingChairs val | EPE: 0.83 vs 0.78 (GMA 0.80) | Hurt |
| enlarge GRU kernel instead of SKBlock (ablation) | GMA updater · Sintel train Clean (C+T) | EPE: SKMoE+SKBlock 1.22 vs LargeGRU 1.32 | Hurt |
Closest papers by meaning
- Recurrent Partial Kernel Network for Efficient Optical Flow Estimation (AAAI 2024 · similarity 0.91)
- SAMFlow: Eliminating Any Fragmentation in Optical Flow with Segment Anything Model (AAAI 2024 · similarity 0.88)
- DPFlow: Adaptive Optical Flow Estimation with a Dual-Pyramid Framework (CVPR 2025 · similarity 0.88)
- Taming generative video models for zero-shot optical flow extraction (NeurIPS 2025 · similarity 0.88)
- UnSAMFlow: Unsupervised Optical Flow Guided by Segment Anything Model (CVPR 2024 · similarity 0.88)
- StreamFlow: Streamlined Multi-Frame Optical Flow Estimation for Video Sequences (NeurIPS 2024 · similarity 0.87)
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.