MyeongJin-Kim / Learning-Texture-Invariant-Representation

Learning Texture Invariant Representation for Domain Adaptation of Semantic Segmentation, CVPR 2020

Learning Texture Invariant Representation for Domain Adaptation of Semantic Segmentation

A pytorch implementation of LTIR.

Requirements

• Python 3.6
• torch==1.2
• torchvision==0.4
• Pillow==6.1.0

Preparing dataset

We used code from Style-swap and CycleGAN.

Training

Initial weight

• ResNet

• VGG

python train_gta2cityscapes.py --translated-data-dir /Path/to/translated/source --stylized-data-dir /Path/to/stylized/source

Evalutation

python evaluate_cityscapes.py --restore-from /Path/to/weight
python compute_iou.py /Path/to/Cityscapes/gtFine/val /Path/to/results

Weight of Final Model

GTA5 to Cityscapes

• ResNet
• VGG

SYNTHIA to Cityscapes

• ResNet
• VGG

Acknowledgement

This code is based on AdaptSegNet and BDL.