This repository contains information to reproduce our experiments.
@inproceedings{kalb2023featurereuse,
title={Principles of Forgetting in Domain-Incremental Semantic Segmentation in Adverse Weather Conditions},
authors={Kalb, Tobias and Beyerer, J\"urgen},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2023}
}
The data augmentation pipelines we used are described in the yaml-Files. These include:
- basic augmentation for Cityscapes
- basic augmentations used for training on the ACDC subsets
- Distortion
- Gaus
- Noise
- AutoAlbum with the corresonoding search configuration and the resulting augmentation config.
As they directly contain our used arguments for the Albumentations transformations, they can also be directly loaded as composed transformations, as shown in ourr sample notebook.
The checkpoints for ResNet50 trained with various SSL methods can be found in their respective repositories:
For ErfNet trained with MoCo v3 and DINO the checkpoints can be found here: https://drive.google.com/drive/folders/1HH4F7gMm4FxqyZcgNBuwicrCFb0r40mh?usp=sharing.
The segmentation models and the corresponding ImageNet weights were based on the following implementations:
- DeepLabV3+ from github.com/qubvel/segmentation_models.pytorch (decoder_channels=512, encoder_output_stride= 8, encoder_name= "resnet50")
- ERFNet from github.com/Eromera/erfnet_pytorch
- SegFormer-B2 from HuggingFace
- RTFormer-Base from PaddleSeg
- BiSeNetV2 from github.com/CoinCheung/BiSeNet
- SegHRNet-w48 from github.com/HRNet/HRNet-Semantic-Segmentation