Research map / 3D vision / Depth and stereo
What Makes Good Synthetic Training Data for Learning Disparity and Optical Flow Estimation?
Nikolaus Mayer, Eddy Ilg, Philipp Fischer, Caner Hazırbaş, Daniel Cremers, Alexey Dosovitskiy, Thomas Brox
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 |
|---|---|---|---|
| add scaling motion to training data (rotation only -> with scaling) (ablation) | FlowNetC on synthetic 2D shapes · KITTI 2015 | EPE: 18.07 -> 15.48 | Improved |
| add holes in objects to training data (ablation) | FlowNetC on synthetic 2D shapes · Sintel train clean | EPE: 4.52 -> 4.71 | Hurt |
| add non-rigid deformations to training data (ablation) | FlowNetC on synthetic 2D shapes · KITTI 2015 | EPE: 16.45 -> 14.97 | Improved |
| add rotation motion to training data (ablation) | FlowNetC on synthetic 2D shapes · KITTI 2015 | EPE: 17.90 -> 18.07 | Hurt |
| use real Flickr photographs as textures instead of Plasma textures (ablation) | FlowNetC · Sintel | EPE: 5.85 -> 4.57 | Improved |
| match training displacement distribution to test (Sintel-like vs 3x Sintel-like) (ablation) | FlowNetC · Sintel | EPE: 5.75 -> 4.60 | Improved |
| train with static lighting instead of shadeless (ablation) | FlowNetC on 3D-remodeled FlyingChairs · Sintel | EPE: 4.41 -> 4.16 | Improved |
| train with dynamic lighting instead of static (ablation) | FlowNetC on 3D-remodeled FlyingChairs · Sintel | EPE: 4.16 -> 4.22 | No clear effect |
| 1:1 static/dynamic lighting mixture (ablation) | FlowNetC on 3D-remodeled FlyingChairs · Sintel | EPE: 4.24 | No clear effect |
| add color and geometry augmentation (no augmentation -> full augmentation) (ablation) | FlowNetC · Sintel | EPE: 7.66 -> 4.60 | Improved |
| add geometry augmentation between frames on top of augmentation on both frames (ablation) | FlowNetC · Sintel | EPE: 5.25 -> 5.33 | Hurt |
| add color changes between frames (ablation) | FlowNetC · Sintel | EPE: 6.30 -> 6.12 | Improved |
Closest papers by meaning
- A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation (CVPR 2016 · similarity 0.92)
- Learning Optical Flow From Still Images (CVPR 2021 · similarity 0.90)
- Flow-Anything: Learning Real-World Optical Flow Estimation From Large-Scale Single-View Images (TPAMI 2025 · similarity 0.88)
- AutoFlow: Learning a Better Training Set for Optical Flow (CVPR 2021 · similarity 0.88)
- Investigating Synthetic-to-Real Transfer Robustness for Stereo Matching and Optical Flow Estimation (TPAMI 2025 · similarity 0.88)
- Taming generative video models for zero-shot optical flow extraction (NeurIPS 2025 · 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.
