Omnisupervised Omnidirectional Semantic Segmentation
PASS Datast Panoramic Annular Semantic Segmentation Dataset with pixel-wise labels (400 images).
Chengyuan Dataset Panoramas captured with an instrumented vehicle (650 images).
Streetview Dataset Panoramas collected in different cities including New York, Beijing, Shanghai, Changsha, Hangzhou, Huddersfield, Madrid, Karlsruhe and Sydney.
Training:
CUDA_VISIBLE_DEVICES=0,1,2,3
python3 segment.py
--basedir /home/kyang/Downloads/
--num-epochs 200
--batch-size 12
--savedir /erfpsp
--datasets 'MAP' 'IDD20K'
--num-samples 18000
--alpha 0
--beta 0
--model erfnet_pspnet
Evaluation:
python3 eval_color.py
--datadir /home/kyang/Downloads/Mapillary/
--subset val
--loadDir ./trained/
--loadWeights model_best.pth
--loadModel erfnet_pspnet.py
--basedir /home/kyang/Downloads/
--datasets 'MAP' 'IDD20K'
If you use our code or dataset, please consider referencing the following paper:
Omnisupervised Omnidirectional Semantic Segmentation. K. Yang, X. Hu, Y. Fang, K. Wang, R. Stiefelhagen. IEEE Transactions on Intelligent Transportation Systems (T-ITS), September 2020. [PDF]
@article{yang2020omnisupervised,
title={Omnisupervised Omnidirectional Semantic Segmentation},
author={Yang, Kailun and Hu, Xinxin and Fang, Yicheng and Wang, Kaiwei and Stiefelhagen, Rainer},
journal={IEEE Transactions on Intelligent Transportation Systems},
year={2020},
publisher={IEEE}
}