Method Details


Details for method 'LGE A&B Center: HANet (ResNet-101)'

 

Method overview

name LGE A&B Center: HANet (ResNet-101)
challenge pixel-level semantic labeling
details Dataset: "fine train + fine val", No coarse, Backbone: ImageNet pretrained ResNet-101
publication Cars Can’t Fly up in the Sky: Improving Urban-Scene Segmentation via Height-driven Attention Networks
Sungha Choi (LGE, Korea Univ.), Joanne T. Kim (Korea Univ.), Jaegul Choo (KAIST)
CVPR 2020
https://arxiv.org/abs/2003.05128
project page / code https://github.com/shachoi/HANet
used Cityscapes data fine annotations
used external data ImageNet
runtime n/a
subsampling no
submission date January, 2020
previous submissions

 

Average results

Metric Value
IoU Classes 82.0923
iIoU Classes 62.1944
IoU Categories 91.9537
iIoU Categories 81.157

 

Class results

Class IoU iIoU
road 98.8433 -
sidewalk 88.0083 -
building 93.9228 -
wall 60.526 -
fence 63.3308 -
pole 71.2508 -
traffic light 78.1123 -
traffic sign 81.3129 -
vegetation 94.0438 -
terrain 72.8905 -
sky 96.0918 -
person 87.8972 71.3994
rider 74.4754 54.7371
car 96.4582 91.6049
truck 76.9925 46.1673
bus 88.0295 60.5
train 85.8806 57.0095
motorcycle 72.6796 50.5946
bicycle 79.0076 65.5425

 

Category results

Category IoU iIoU
flat 98.7779 -
nature 93.7127 -
object 76.7702 -
sky 96.0918 -
construction 94.2231 -
human 88.0922 72.4296
vehicle 96.0079 89.8844

 

Links

Download results as .csv file

Benchmark page