Panoramic Semantic Segmentation in the Wild
WildPASS Dataset with City Names
The WildPASS dataset contains annotated 500 panoramas taken from 25 cities located on multiple continents for evaludation.
The WildPASS2K dataset contains 2000 unlabled panoramas taken from 40 cities, which can be used for facilitating domain adapation and creating pseudo labels.
For training, we suggest to use Mapillary Vistas, or with a combination of IDD20K, Cityscapes, ApolloScape, BDD10K, Audi A2D2, KITTI, KITTI-360, and WildDash2 datasets.
Example
CUDA_VISIBLE_DEVICES=0
python3 eval_color_fusion.py
--datasets 'MAP' 'IDD20K'
--is-fuse
--basedir /cvhci/data/
--subset val
--loadDir ../trained_models/
--loadWeights model_best.pth
--loadModel ecanet.py
--datadir /cvhci/data/WildPASS
If you use our code or dataset, please consider citing any of the following papers:
Capturing Omni-Range Context for Omnidirectional Segmentation. K. Yang, J. Zhang, S. Reiß, X. Hu, R. Stiefelhagen. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, United States (Virtual), June 2021. [PDF]
@inproceedings{yang2021capturing,
title={Capturing Omni-Range Context for Omnidirectional Segmentation},
author={Yang, Kailun and Zhang, Jiaming and Rei{\ss}, Simon and Hu, Xinxin and Stiefelhagen, Rainer},
booktitle={2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2021}
}
Is Context-Aware CNN Ready for the Surroundings? Panoramic Semantic Segmentation in the Wild. K. Yang, X. Hu, R. Stiefelhagen. IEEE Transactions on Image Processing (TIP), 2021. [PDF]
@article{yang2021context,
title={Is Context-Aware CNN Ready for the Surroundings? Panoramic Semantic Segmentation in the Wild},
author={Yang, Kailun and Hu, Xinxin and Stiefelhagen, Rainer},
journal={IEEE Transactions on Image Processing},
volume={30},
pages={1866--1881},
year={2021},
publisher={IEEE}
}