In this paper, we propose a novel method in pursuit of solving the anomaly segmentation task. Based on previous generative approaches to anomaly segmentation, we follow the fashion of modeling likelihood
We suggest downloading datasets under a new directory data/
.
This implementation is built on MMSegmentation v0.22.1. Many thanks to the contributors for their great efforts.
Please follow the get_started for installation and dataset_prepare for dataset preparation.
Other requirements: pip install timm==0.5.4 einops==0.4.1
# multi-gpu train
bash tools/dist_train.sh configs/_gmmseg/segformer_mit-b5_gmmseg_1024x2048_160k_cityscapes.py $GPU_NUM
# single-gpu test
python tools/test_fishyscapes.py configs/_gmmseg/segformer_mit-b5_gmmseg_fishyscapes.py /path/to/checkpoint_file
Our code implementation is based on GMMSeg, and we would like to thank the authors for their comprehensive work.