Research map / Architectures and efficiency / RNNs and normalization
FlowFormer++: Masked Cost Volume Autoencoding for Pretraining Optical Flow Estimation
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 |
|---|---|---|---|
| Pretrain FlowFormer cost-volume encoder with MCVA | FlowFormer · Sintel train, C+T | AEPE: 0.94 -> 0.90 (Clean); 2.33 -> 2.30 (Final) | Helped |
| Pretrain FlowFormer cost-volume encoder with MCVA | FlowFormer · KITTI-15 train, C+T | F1-all: 14.72 -> 14.13 | Helped |
| Pretrain FlowFormer cost-volume encoder with MCVA | FlowFormer · Sintel test and KITTI-15 test, C+T+S+K+H | AEPE / F1-all: 1.16 -> 1.07 (Sintel Clean); 2.09 -> 1.94 (Sintel Final); 4.68 -> 4.52 (KITTI-15 F1-all) | Helped |
| Use block-sharing masking rather than random masking (ablation) | MCVA-pretrained FlowFormer · Sintel and KITTI-15 train, C+T | AEPE / F1-epe / F1-all: 0.93 -> 0.90 (Clean); 2.35 -> 2.30 (Final); 4.03 -> 3.93 (F1-epe); 14.30 -> 14.13 (F1-all) | Mixed |
| Set masking ratio to 50% rather than 20% or 80% (ablation) | MCVA-pretrained FlowFormer · Sintel and KITTI-15 train, C+T | AEPE / F1-epe / F1-all: 0.97 -> 0.90 -> 0.94 (Clean); 2.41 -> 2.30 -> 2.35 (Final); 3.91 -> 3.93 -> 4.00 (F1-epe); 14.27 -> 14.13 -> 14.07 (F1-a | Mixed |
| Use random rather than fixed reconstruction locations (ablation) | MCVA-pretrained FlowFormer · Sintel and KITTI-15 train, C+T | AEPE / F1-epe / F1-all: 0.99 -> 0.95 (Clean); 2.40 -> 2.42 (Final); 4.35 -> 3.99 (F1-epe); 15.33 -> 14.47 (F1-all) | Helped |
| Add local cropped cost-patch query features to random-location reconstruction (ablation) | MCVA-pretrained FlowFormer · Sintel and KITTI-15 train, C+T | AEPE / F1-epe / F1-all: 0.95 -> 0.90 (Clean); 2.42 -> 2.30 (Final); 3.99 -> 3.93 (F1-epe); 14.47 -> 14.13 (F1-all) | Helped |
| Freeze the context encoder during pretraining (ablation) | MCVA-pretrained FlowFormer · Sintel and KITTI-15 train, C+T | AEPE / F1-epe / F1-all: 0.92 -> 0.90 (Clean); 2.34 -> 2.30 (Final); 3.90 -> 3.93 (F1-epe); 14.23 -> 14.13 (F1-all) | Mixed |
| Freeze the image encoder during pretraining (ablation) | MCVA-pretrained FlowFormer · Pretraining stability | convergence | Helped |
| Use MCVA rather than photometric and smoothness-loss pretraining (ablation) | FlowFormer · Sintel train, C+T | AEPE: 0.99 -> 0.90 (Clean); 2.54 -> 2.30 (Final) | Helped |
Closest papers by meaning
- SAMFlow: Eliminating Any Fragmentation in Optical Flow with Segment Anything Model (AAAI 2024 · similarity 0.86)
- DPFlow: Adaptive Optical Flow Estimation with a Dual-Pyramid Framework (CVPR 2025 · similarity 0.85)
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · similarity 0.85)
- Extending Video Masked Autoencoders to 128 frames (NeurIPS 2024 · similarity 0.85)
- FlowVid: Taming Imperfect Optical Flows for Consistent Video-to-Video Synthesis (CVPR 2024 · similarity 0.84)
- StreamFlow: Streamlined Multi-Frame Optical Flow Estimation for Video Sequences (NeurIPS 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.