GSTO: Gated Scale-Transfer Operation for Multi-Scale Feature Learning in Pixel Labeling

This is the official implementation of paper: GSTO: Gated Scale-Transfer Operation for Multi-Scale Feature Learning in Pixel Labeling

by Zhuoying Wang, Yongtao Wang.

Introduction

The project is only free for academic research purposes, but needs authorization for commerce. For commerce permission, please contact wyt@pku.edu.cn.

Citation

If you use our code/model/data, please cite our paper: https://arxiv.org/abs/2005.13363

Semantic Segmentation Results

Cityscapes val set

single scale and no flipping, without using OHEM

model	#Params	GFLOPs	mIoU
GSTO-HRNet-W48	65.93M	714.0	82.1

Cityscapes test set

multi-scale with flipping

model	Train set	OHEM	mIoU
GSTO-HRNet-W48	Train	False	81.9
GSTO-HRNet-W48	Trainval	False	82.3
GSTO-HRNet-W48	Trainval	True	82.4

LIP

model	Extra.	Pixel acc.	avg acc.	mIoU
GSTO-HRNet-W48	Train	88.38	68.36	57.4

Pose Estimation Results

COCO val set

model	Input size	#Params	GFLOPs	AP
GSTO-HRNet-W32	384x288	29.6M	18.2	76.5
GSTO-HRNet-W48	384x288	66.0M	37.6	76.7

COCO test set

model	Input size	#Params	GFLOPs	AP
GSTO-HRNet-W32	384x288	29.6M	18.2	75.5
GSTO-HRNet-W48	384x288	66.0M	37.6	75.8

Quick inference

Install

PyTorch=0.4.1
git clone https://github.com/VDIGPKU/GSTO
install dependencies EasyDict==1.7 opencv-python==3.4.1.15 shapely==1.6.4 Cython scipy pandas pyyaml json_tricks scikit-image yacs>=0.1.5 tensorboardX>=1.6 tqdm ninja

Data

follow the prepare instruction in HRNet-Semantic-Segmentation

Test

cityscapes test mIoU 82.4

CUDA_VISIBLE_DEVICES=0 python test.py --cfg cfg_files/gsto_hrnet_w48_cityscapes.yaml

LIP test mIoU 57.4

CUDA_VISIBLE_DEVICES=0 python test.py --cfg cfg_files/gsto_hrnet_w48_lip.yaml

VDIGPKU/GSTO