# You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection

**Authors:** Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, Wenyu Liu  
**arXiv:** [2106.00666](https://arxiv.org/abs/2106.00666)  
**Date:** 2021-06-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.00666v3/

## Key points

- Authors explore the capability of Transformers in 2D object- and region-level recognition without prior knowledge about spatial structure
- They propose a series of object detection models called YOLOS based on vanilla Vision Transformer with minimal modifications
- YOLOS achieves competitive results on COCO object detection benchmark, even without extensive pre-training or architectural changes
- YOLOS-Base achieves a box Average Precision (AP) of 42.0 on COCO val
- The paper discusses impacts and limitations of current pre-training schemes and model scaling strategies for Transformers in vision tasks
- Insights provided on how these factors affect performance and suggestions for future improvement
- Transformers can effectively handle 2D object- and region-level recognition tasks with minimal knowledge about spatial structure
- Potential of Transformers in computer vision applications highlighted

## Summary

In the paper "You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection," authors Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu and Wenyu Liu explore the capability of Transformers to perform 2D object- and region-level recognition without prior knowledge about the 2D spatial structure. They propose a series of object detection models called You Only Look at One Sequence (YOLOS), which are based on the vanilla Vision Transformer with minimal modifications and inductive biases specific to the target task. The authors conduct experiments using YOLOS pre-trained on the mid-sized ImageNet-1k dataset and evaluate its performance on the challenging COCO object detection benchmark. Surprisingly, even without extensive pre-training or architectural changes, YOLOS achieves competitive results. For instance, YOLOS-Base directly adopted from BERT-Base architecture achieves a box Average Precision (AP) of 42.0 on COCO val. The paper also discusses the impacts and limitations of current pre-training schemes and model scaling strategies for Transformers in vision tasks through the lens of YOLOS. The authors provide insights into how these factors affect performance and suggest future directions for improvement. Overall, this study demonstrates that Transformers can effectively handle 2D object- and region-level recognition tasks with minimal knowledge about spatial structure. The findings highlight the potential of Transformers in computer vision applications and provide valuable insights for further research in this area. Code implementations and pre-trained models are available at https://github.com/hustvl/YOLOS.