Research map / Architectures and efficiency / Attention and transformers
UniFormer: Unifying Convolution and Self-Attention for Visual Recognition
Kunchang Li, Yali Wang, Junhao Zhang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, Yu Qiao
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 Token Labeling (ablation) | UniFormer-S · ImageNet-1K val | top-1 accuracy: 82.9 -> 83.4 | Improved |
| add Token Labeling (ablation) | UniFormer-B · ImageNet-1K val | top-1 accuracy: 83.9 -> 85.1 | Improved |
| fine-tune at 384x384 resolution | UniFormer-S with Token Labeling · ImageNet-1K val | top-1 accuracy: 83.4 -> 84.6 | Improved |
| use overlapped patch embedding and more blocks (ablation) | UniFormer-S · ImageNet-1K val | top-1 accuracy: 82.9 -> 83.4 | Improved |
| use global MHRA instead of window-based hybrid in Stage3 (ablation) | UniFormer-S · ADE20K Semantic FPN | mIoU: 46.2 -> 46.6 mIoU but 199 -> 247 GFLOPs | Improved |
| use global MHRA instead of window-based hybrid in Stage3 (ablation) | UniFormer-B · ADE20K Semantic FPN | mIoU: 47.7 -> 48.0 | Improved |
| use global MHRA instead of window-based hybrid in Stage3 (ablation) | UniFormer-S · ADE20K UperNet | mIoU: 47.0 -> 47.6 | Improved |
| pretrain on Kinetics-600 instead of Kinetics-400 | UniFormer-S 16 frames · Something-Something V1 | top-1 accuracy: 53.8 -> 54.4 | Improved |
| pretrain on Kinetics-600 instead of Kinetics-400 | UniFormer-B 16 frames · Something-Something V1 | top-1 accuracy: 59.1 -> 58.8 | No clear effect |
| increase from 16 to 32 frames | UniFormer-B · Kinetics-400 | top-1 accuracy: 82.0 -> 82.9 | Improved |
| increase from 1 to 3 spatial crops | UniFormer-B 32 frames · Kinetics-400 | top-1 accuracy: 82.9 -> 83.0 | No clear effect |
Closest papers by meaning
- Bilinear Convolutional Neural Networks for Fine-Grained Visual Recognition (TPAMI 2017 · similarity 0.91)
- Involution: Inverting the Inherence of Convolution for Visual Recognition (CVPR 2021 · similarity 0.91)
- UniFormer: Unified Transformer for Efficient Spatial-Temporal Representation Learning (ICLR 2022 · similarity 0.91)
- CvT: Introducing Convolutions to Vision Transformers (ICCV 2021 · similarity 0.90)
- SparseFormer: Sparse Visual Recognition via Limited Latent Tokens (ICLR 2024 · similarity 0.90)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.89)
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.
