pspnet

PSPNet

Pyramid Scene Parsing Network

Introduction

Official Repo

Code Snippet

Abstract

Scene parsing is challenging for unrestricted open vocabulary and diverse scenes. In this paper, we exploit the capability of global context information by different-region-based context aggregation through our pyramid pooling module together with the proposed pyramid scene parsing network (PSPNet). Our global prior representation is effective to produce good quality results on the scene parsing task, while PSPNet provides a superior framework for pixel-level prediction tasks. The proposed approach achieves state-of-the-art performance on various datasets. It came first in ImageNet scene parsing challenge 2016, PASCAL VOC 2012 benchmark and Cityscapes benchmark. A single PSPNet yields new record of mIoU accuracy 85.4% on PASCAL VOC 2012 and accuracy 80.2% on Cityscapes.

PSPNet-R50 D8 model structure

Results and models

Cityscapes

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
PSPNet	R-50-D8	512x1024	40000	6.1	4.07	V100	77.85	79.18	config	model | log
PSPNet	R-101-D8	512x1024	40000	9.6	2.68	V100	78.34	79.74	config	model | log
PSPNet	R-50-D8	769x769	40000	6.9	1.76	V100	78.26	79.88	config	model | log
PSPNet	R-101-D8	769x769	40000	10.9	1.15	V100	79.08	80.28	config	model | log
PSPNet	R-18-D8	512x1024	80000	1.7	15.71	V100	74.87	76.04	config	model | log
PSPNet	R-50-D8	512x1024	80000	-	-	V100	78.55	79.79	config	model | log
PSPNet	R-50b-D8 rsb	512x1024	80000	6.2	3.82	V100	78.47	79.45	config	model | log
PSPNet	R-101-D8	512x1024	80000	-	-	V100	79.76	81.01	config	model | log
PSPNet (FP16)	R-101-D8	512x1024	80000	5.34	8.77	V100	79.46	-	config	model | log
PSPNet	R-18-D8	769x769	80000	1.9	6.20	V100	75.90	77.86	config	model | log
PSPNet	R-50-D8	769x769	80000	-	-	V100	79.59	80.69	config	model | log
PSPNet	R-101-D8	769x769	80000	-	-	V100	79.77	81.06	config	model | log
PSPNet	R-18b-D8	512x1024	80000	1.5	16.28	V100	74.23	75.79	config	model | log
PSPNet	R-50b-D8	512x1024	80000	6.0	4.30	V100	78.22	79.46	config	model | log
PSPNet	R-101b-D8	512x1024	80000	9.5	2.76	V100	79.69	80.79	config	model | log
PSPNet	R-18b-D8	769x769	80000	1.7	6.41	V100	74.92	76.90	config	model | log
PSPNet	R-50b-D8	769x769	80000	6.8	1.88	V100	78.50	79.96	config	model | log
PSPNet	R-101b-D8	769x769	80000	10.8	1.17	V100	78.87	80.04	config	model | log
PSPNet	R-50-D32	512x1024	80000	3.0	15.21	V100	73.88	76.85	config	model | log
PSPNet	R-50b-D32 rsb	512x1024	80000	3.1	16.08	V100	74.09	77.18	config	model | log
PSPNet	R-50b-D32	512x1024	80000	2.9	15.41	V100	72.61	75.51	config	model | log

ADE20K

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
PSPNet	R-50-D8	512x512	80000	8.5	23.53	V100	41.13	41.94	config	model | log
PSPNet	R-101-D8	512x512	80000	12	15.30	V100	43.57	44.35	config	model | log
PSPNet	R-50-D8	512x512	160000	-	-	V100	42.48	43.44	config	model | log
PSPNet	R-101-D8	512x512	160000	-	-	V100	44.39	45.35	config	model | log

Pascal VOC 2012 + Aug

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
PSPNet	R-50-D8	512x512	20000	6.1	23.59	V100	76.78	77.61	config	model | log
PSPNet	R-101-D8	512x512	20000	9.6	15.02	V100	78.47	79.25	config	model | log
PSPNet	R-50-D8	512x512	40000	-	-	V100	77.29	78.48	config	model | log
PSPNet	R-101-D8	512x512	40000	-	-	V100	78.52	79.57	config	model | log

Pascal Context

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
PSPNet	R-101-D8	480x480	40000	8.8	9.68	V100	46.60	47.78	config	model | log
PSPNet	R-101-D8	480x480	80000	-	-	V100	46.03	47.15	config	model | log

Pascal Context 59

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
PSPNet	R-101-D8	480x480	40000	-	-	V100	52.02	53.54	config	model | log
PSPNet	R-101-D8	480x480	80000	-	-	V100	52.47	53.99	config	model | log

Dark Zurich and Nighttime Driving

We support evaluation results on these two datasets using models above trained on Cityscapes training set.

Method	Backbone	Training Dataset	Test Dataset	mIoU	config	evaluation checkpoint
PSPNet	R-50-D8	Cityscapes Training set	Dark Zurich	10.91	config	model | log
PSPNet	R-50-D8	Cityscapes Training set	Nighttime Driving	23.02	config	model | log
PSPNet	R-50-D8	Cityscapes Training set	Cityscapes Validation set	77.85	config	model | log
PSPNet	R-101-D8	Cityscapes Training set	Dark Zurich	10.16	config	model | log
PSPNet	R-101-D8	Cityscapes Training set	Nighttime Driving	20.25	config	model | log
PSPNet	R-101-D8	Cityscapes Training set	Cityscapes Validation set	78.34	config	model | log
PSPNet	R-101b-D8	Cityscapes Training set	Dark Zurich	15.54	config	model | log
PSPNet	R-101b-D8	Cityscapes Training set	Nighttime Driving	22.25	config	model | log
PSPNet	R-101b-D8	Cityscapes Training set	Cityscapes Validation set	79.69	config	model | log

COCO-Stuff 10k

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
PSPNet	R-50-D8	512x512	20000	9.6	20.5	V100	35.69	36.62	config	model | log
PSPNet	R-101-D8	512x512	20000	13.2	11.1	V100	37.26	38.52	config	model | log
PSPNet	R-50-D8	512x512	40000	-	-	V100	36.33	37.24	config	model | log
PSPNet	R-101-D8	512x512	40000	-	-	V100	37.76	38.86	config	model | log

COCO-Stuff 164k

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
PSPNet	R-50-D8	512x512	80000	9.6	20.5	V100	38.80	39.19	config	model | log
PSPNet	R-101-D8	512x512	80000	13.2	11.1	V100	40.34	40.79	config	model | log
PSPNet	R-50-D8	512x512	160000	-	-	V100	39.64	39.97	config	model | log
PSPNet	R-101-D8	512x512	160000	-	-	V100	41.28	41.66	config	model | log
PSPNet	R-50-D8	512x512	320000	-	-	V100	40.53	40.75	config	model | log
PSPNet	R-101-D8	512x512	320000	-	-	V100	41.95	42.42	config	model | log

LoveDA

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
PSPNet	R-18-D8	512x512	80000	1.45	26.87	V100	48.62	47.57	config	model | log
PSPNet	R-50-D8	512x512	80000	6.14	6.60	V100	50.46	50.19	config	model | log
PSPNet	R-101-D8	512x512	80000	9.61	4.58	V100	51.86	51.34	config	model | log

Potsdam

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
PSPNet	R-18-D8	512x512	80000	1.50	85.12	V100	77.09	78.30	config	model | log
PSPNet	R-50-D8	512x512	80000	6.14	30.21	V100	78.12	78.98	config	model | log
PSPNet	R-101-D8	512x512	80000	9.61	19.40	V100	78.62	79.47	config	model | log

Vaihingen

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
PSPNet	R-18-D8	512x512	80000	1.45	85.06	V100	71.46	73.36	config	model | log
PSPNet	R-50-D8	512x512	80000	6.14	30.29	V100	72.36	73.75	config	model | log
PSPNet	R-101-D8	512x512	80000	9.61	19.97	V100	72.61	74.18	config	model | log

iSAID

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
PSPNet	R-18-D8	896x896	80000	4.52	26.91	V100	60.22	61.25	config	model | log
PSPNet	R-50-D8	896x896	80000	16.58	8.88	V100	65.36	66.48	config	model | log

Note:

• FP16 means Mixed Precision (FP16) is adopted in training.
• 896x896 is the Crop Size of iSAID dataset, which is followed by the implementation of PointFlow: Flowing Semantics Through Points for Aerial Image Segmentation
• rsb is short for 'Resnet strikes back'.
• The b in R-50b means ResNetV1b, which is a standard ResNet backbone. In MMSegmentation, default backbone is ResNetV1c, which usually performs better in semantic segmentation task.

Citation

@inproceedings{zhao2017pspnet,
  title={Pyramid Scene Parsing Network},
  author={Zhao, Hengshuang and Shi, Jianping and Qi, Xiaojuan and Wang, Xiaogang and Jia, Jiaya},
  booktitle={CVPR},
  year={2017}
}

@article{wightman2021resnet,
  title={Resnet strikes back: An improved training procedure in timm},
  author={Wightman, Ross and Touvron, Hugo and J{\'e}gou, Herv{\'e}},
  journal={arXiv preprint arXiv:2110.00476},
  year={2021}
}