Research map / Video / Video segmentation and tracking
TransFlow: Transformer As Flow Learner
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 |
|---|---|---|---|
| adding cross-attention to self-attention in the encoder (ablation) | TransFlow encoder · Sintel (val) / KITTI (val) | AEPE clean / F1-all: Sintel clean 0.58->0.55; KITTI F1-all 1.22->1.18 | Helped |
| adding temporal association across multiple frames (ablation) | TransFlow encoder · Sintel (val) / KITTI (val) | AEPE clean / F1-all: Sintel clean 0.55->0.49; KITTI F1-all 1.18 unchanged | Helped |
| adding strategic (learned score-map) masking in self-supervised pre-training (ablation) | TransFlow · Sintel (val) / KITTI (val) | AEPE clean / F1-all: Sintel clean 0.49->0.44; KITTI F1-all 1.18->1.07 | Helped |
| adding occlusion consistency loss (ablation) | TransFlow · Sintel (val) / KITTI (val) | AEPE clean / F1-all: Sintel clean 0.44->0.42; KITTI F1-all 1.07->1.05 | Helped |
| increase patch size from 8x8 to 16x16 (ablation) | TransFlow · Sintel (val) / KITTI (val) | AEPE clean / F1-all: Sintel clean 0.42->0.59; KITTI F1-all 1.05->1.21 | Hurt |
| learned (strategic) masking vs random masking (ablation) | TransFlow self-supervised pre-training · Sintel (val) | AEPE clean: 0.50 -> 0.42 | Helped |
| masking ratio of 90% instead of 50% (ablation) | TransFlow self-supervised pre-training · Sintel (val) | AEPE clean: 0.42 -> 0.49 | Hurt |
| relative instead of learnable absolute positional embedding (ablation) | TransFlow · Sintel (val) | AEPE clean: 0.42 -> 0.47 | Hurt |
| adding TransFlow encoder features to RDN, SELSA, PTSEFormer video object detectors | RDN, SELSA, PTSEFormer · ImageNet VID | mAP (%): +3.7, +2.6, +1.7 | Helped |
| adding TransFlow encoder features to SuperSloMo and IFR-Net interpolators | SuperSloMo, IFR-Net · GoPro video interpolation | PSNR / SSIM: PSNR +0.29 and +0.18 dB; SSIM +0.014 and +0.012 | Helped |
| adding TransFlow encoder features to StabNet and PWStableNet | StabNet, PWStableNet · DeepStab video stabilization | Distortion / Stability: StabNet D +0.02, S +0.04; PWStableNet D +0.03, S +0.02 | Mixed |
| self-supervised pre-training on target data (no multi-stage pre-training) with TransFlow | TransFlow · Sintel test clean/final, KITTI-15 test | AEPE / F1-all: Sintel clean 1.06, final 2.08; KITTI F1-all 4.32 vs Flowformer 4.68 | Helped |
Closest papers by meaning
- CRAFT: Cross-Attentional Flow Transformer for Robust Optical Flow (CVPR 2022 · similarity 0.89)
- StreamFlow: Streamlined Multi-Frame Optical Flow Estimation for Video Sequences (NeurIPS 2024 · similarity 0.87)
- SAMFlow: Eliminating Any Fragmentation in Optical Flow with Segment Anything Model (AAAI 2024 · similarity 0.87)
- Recurrent Partial Kernel Network for Efficient Optical Flow Estimation (AAAI 2024 · similarity 0.86)
- DPFlow: Adaptive Optical Flow Estimation with a Dual-Pyramid Framework (CVPR 2025 · similarity 0.86)
- Taming generative video models for zero-shot optical flow extraction (NeurIPS 2025 · similarity 0.86)
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.