Compound U-Net: Multiscale Feature Extraction Benefits Solar Filament Segmentation and Detection 1.Instruction We used python 3.9 and conda 24.11.3 on a single RTX A6000 GPU.install required packages: pip install -r segmentation/requirements.txt if on linux server, PYTHONPATH may need to be set. export PYTHONPATH=$PYTHONPATH:{your_path_to_filament_dir} for training, run: python segmentation/train.py for inference, run: python segmentation/inference.py for clustering, run: python segmentation/postprocessing.py 2.For model modification The Compound U-Net can be modified in segmentation/utils/models/backbone/compound_unet.py, and the output module can be modified in segmentation/utils/models/filament_seg.py 3.Introduction and access to the datasets Each dataset is made up of the training set with 80 images, the validation set with 20 images and the test setwith 20 images.The HAS dataset is from Zheng et al., which is available at DOI:10.5281/zenodo.10598419.The MHAS dataset is manually corrected by ourselves after the automatic annotation proposed by Zheng et al.,which is available at DOI:10.5281/zenodo.17230605.The annotations are all single-channel images with only two values. In HAS, the solar filaments are labeled as 1.In MHAS, the solar filaments are labeled as 255. 4.Performance of our model Model Dataset Test IoU Test F1-Score Compound U-Net MHAS 0.677 0.803 Compound U-Net HAS 0.714 0.831