Method Details


Details for method 'GALD-Net'

 

Method overview

name GALD-Net
challenge pixel-level semantic labeling
details We propose Global Aggregation then Local Distribution (GALD) scheme to distribute global information to each position adaptively according to the local information around the position. (Joint work: Key Laboratory of Machine Perception, School of EECS @Peking University and DeepMotion AI Research )
publication Global Aggregation then Local Distribution for Scene Understanding
Xiangtai Li, Ansheng You, Maoke Yang, Kuiyuan Yang, Yunhai Tong
project page / code https://github.com/lxtGH/GALD-Net
used Cityscapes data fine annotations, coarse annotations, 16bit, stereo
used external data Mapillary
runtime n/a
subsampling no
submission date March, 2019
previous submissions

 

Average results

Metric Value
IoU Classes 83.2981
iIoU Classes 64.5173
IoU Categories 92.2831
iIoU Categories 81.9387

 

Class results

Class IoU iIoU
road 98.8097 -
sidewalk 87.6865 -
building 94.2001 -
wall 65.0349 -
fence 66.7147 -
pole 73.1498 -
traffic light 79.2719 -
traffic sign 82.4421 -
vegetation 94.1603 -
terrain 72.9191 -
sky 96.023 -
person 88.4498 73.702
rider 76.233 56.7356
car 96.4998 91.0737
truck 79.821 52.8011
bus 89.5861 59.6299
train 87.6622 60.4326
motorcycle 74.0595 53.6566
bicycle 79.9406 68.1072

 

Category results

Category IoU iIoU
flat 98.7792 -
nature 93.8373 -
object 78.2913 -
sky 96.023 -
construction 94.4293 -
human 88.4802 74.5157
vehicle 96.1418 89.3617

 

Links

Download results as .csv file

Benchmark page