PVT v2: Improved baselines with Pyramid Vision Transformer

Transformers have recently lead to encouraging progress in computer vision. In this work, we present new baselines by improving the original Pyramid Vision Transformer (PVT v1) by adding three designs: (i) a linear complexity attention layer, (ii) an overlapping patch embedding, and (iii) a convolut...

Full description

Saved in:
Bibliographic Details
Published in:Computational visual media (Beijing) Vol. 8; no. 3; pp. 415 - 424
Main Authors: Wang, Wenhai, Xie, Enze, Li, Xiang, Fan, Deng-Ping, Song, Kaitao, Liang, Ding, Lu, Tong, Luo, Ping, Shao, Ling
Format: Journal Article
Language:English
Published: Beijing Tsinghua University Press 01.09.2022
Springer Nature B.V
Subjects:
Archives & records
Artificial Intelligence
Classification
Complexity
Computer Graphics
Computer Science
Computer vision
Image Processing and Computer Vision
Linearity
Research Article
Segmentation
Semantics
User Interfaces and Human Computer Interaction
transformers
object detection
semantic segmentation
image classification
dense prediction
ISSN:2096-0433, 2096-0662
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Transformers have recently lead to encouraging progress in computer vision. In this work, we present new baselines by improving the original Pyramid Vision Transformer (PVT v1) by adding three designs: (i) a linear complexity attention layer, (ii) an overlapping patch embedding, and (iii) a convolutional feed-forward network. With these modifications, PVT v2 reduces the computational complexity of PVT v1 to linearity and provides significant improvements on fundamental vision tasks such as classification, detection, and segmentation. In particular, PVT v2 achieves comparable or better performance than recent work such as the Swin transformer. We hope this work will facilitate state-of-the-art transformer research in computer vision. Code is available at https://github.com/whai362/PVT .
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:2096-0433
2096-0662
DOI:10.1007/s41095-022-0274-8