PVT v2: Improved baselines with pyramid vision transformer

Abstract

Transformer recently has presented encouraging progress in computer vision.\nIn this work, we present new baselines by improving the original Pyramid Vision\nTransformer (PVT v1) by adding three designs, including (1) linear complexity\nattention layer, (2) overlapping patch embedding, and (3) convolutional\nfeed-forward network. With these modifications, PVT v2 reduces the\ncomputational complexity of PVT v1 to linear and achieves significant\nimprovements on fundamental vision tasks such as classification, detection, and\nsegmentation. Notably, the proposed PVT v2 achieves comparable or better\nperformances than recent works such as Swin Transformer. We hope this work will\nfacilitate state-of-the-art Transformer researches in computer vision. Code is\navailable at https://github.com/whai362/PVT.\n

Keywords

Affiliated Institutions

• Inception Institute of Artificial Intelligence AE
• Group Sense (China) HK
• University of Hong Kong HK
• ETH Zurich CH
• Shanghai Artificial Intelligence Laboratory
• Nanjing University of Science and Technology CN

Related Publications