Updates

• (2020/06/21) Code of PVTv2 is released! PVTv2 largely improves PVTv1 and works better than Swin Transformer with ImageNet-1K pre-training.

Pyramid Vision Transformer

The image is from Transformers: Revenge of the Fallen.

This repository contains the official implementation of PVTv1 & PVTv2 in image classification, object detection, and semantic segmentation tasks.

Model Zoo

Image Classification

Classification configs & weights see >>>here<<<.

• PVTv2 on ImageNet-1K

Method	Size	Acc@1	#Params (M)
PVTv2-B0	224	70.5	3.7
PVTv2-B1	224	78.7	14.0
PVTv2-B2-Linear	224	82.1	22.6
PVTv2-B2	224	82.0	25.4
PVTv2-B3	224	83.1	45.2
PVTv2-B4	224	83.6	62.6
PVTv2-B5	224	83.8	82.0

• PVTv1 on ImageNet-1K

Method	Size	Acc@1	#Params (M)
PVT-Tiny	224	75.1	13.2
PVT-Small	224	79.8	24.5
PVT-Medium	224	81.2	44.2
PVT-Large	224	81.7	61.4

Object Detection

Detection configs & weights see >>>here<<<.

• PVTv2 on COCO

Baseline Detectors

Method	Backbone	Pretrain	Lr schd	Aug	box AP	mask AP
RetinaNet	PVTv2-b0	ImageNet-1K	1x	No	37.2	-
RetinaNet	PVTv2-b1	ImageNet-1K	1x	No	41.2	-
RetinaNet	PVTv2-b2	ImageNet-1K	1x	No	44.6	-
RetinaNet	PVTv2-b3	ImageNet-1K	1x	No	45.9	-
RetinaNet	PVTv2-b4	ImageNet-1K	1x	No	46.1	-
RetinaNet	PVTv2-b5	ImageNet-1K	1x	No	46.2	-
Mask R-CNN	PVTv2-b0	ImageNet-1K	1x	No	38.2	36.2
Mask R-CNN	PVTv2-b1	ImageNet-1K	1x	No	41.8	38.8
Mask R-CNN	PVTv2-b2	ImageNet-1K	1x	No	45.3	41.2
Mask R-CNN	PVTv2-b3	ImageNet-1K	1x	No	47.0	42.5
Mask R-CNN	PVTv2-b4	ImageNet-1K	1x	No	47.5	42.7
Mask R-CNN	PVTv2-b5	ImageNet-1K	1x	No	47.4	42.5

Advanced Detectors

Method	Backbone	Pretrain	Lr schd	Aug	box AP	mask AP
Cascade Mask R-CNN	PVTv2-b2-Linear	ImageNet-1K	3x	Yes	50.9	44.0
Cascade Mask R-CNN	PVTv2-b2	ImageNet-1K	3x	Yes	51.1	44.4
ATSS	PVTv2-b2-Linear	ImageNet-1K	3x	Yes	48.9	-
ATSS	PVTv2-b2	ImageNet-1K	3x	Yes	49.9	-
GFL	PVTv2-b2-Linear	ImageNet-1K	3x	Yes	49.2	-
GFL	PVTv2-b2	ImageNet-1K	3x	Yes	50.2	-
Sparse R-CNN	PVTv2-b2-Linear	ImageNet-1K	3x	Yes	48.9	-
Sparse R-CNN	PVTv2-b2	ImageNet-1K	3x	Yes	50.1	-

• PVTv1 on COCO

Detector	Backbone	Pretrain	Lr schd	box AP	mask AP
RetinaNet	PVT-Tiny	ImageNet-1K	1x	36.7	-
RetinaNet	PVT-Small	ImageNet-1K	1x	40.4	-
Mask RCNN	PVT-Tiny	ImageNet-1K	1x	36.7	35.1
Mask RCNN	PVT-Small	ImageNet-1K	1x	40.4	37.8
DETR	PVT-Small	ImageNet-1K	50ep	34.7	-

Semantic Segmentation

Segmentation configs & weights see >>>here<<<.

PVT-v2 + Segmentation see >>>here<<<.

• PVTv1 on ADE20K

Method	Backbone	Pretrain	Iters	mIoU
Semantic FPN	PVT-Tiny	ImageNet-1K	40K	35.7
Semantic FPN	PVT-Small	ImageNet-1K	40K	39.8
Semantic FPN	PVT-Medium	ImageNet-1K	40K	41.6
Semantic FPN	PVT-Large	ImageNet-1K	40K	42.1

License

This repository is released under the Apache 2.0 license as found in the LICENSE file.

Citation

If you use this code for a paper, please cite:

PVTv1

@inproceedings{wang2021pyramid,
  title={Pyramid vision transformer: A versatile backbone for dense prediction without convolutions},
  author={Wang, Wenhai and Xie, Enze and Li, Xiang and Fan, Deng-Ping and Song, Kaitao and Liang, Ding and Lu, Tong and Luo, Ping and Shao, Ling},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={568--578},
  year={2021}
}

PVTv2

@article{wang2021pvtv2,
  title={Pvtv2: Improved baselines with pyramid vision transformer},
  author={Wang, Wenhai and Xie, Enze and Li, Xiang and Fan, Deng-Ping and Song, Kaitao and Liang, Ding and Lu, Tong and Luo, Ping and Shao, Ling},
  journal={Computational Visual Media},
  volume={8},
  number={3},
  pages={1--10},
  year={2022},
  publisher={Springer}
}

Contact

This repo is currently maintained by Wenhai Wang (@whai362), Enze Xie (@xieenze), and Zhe Chen (@czczup).