Official implementation of HATFormer
The SMARS dataset used for evaluation is from A 2D/3D multimodal data simulation approach with applications on urban semantic segmentation, building extraction and change detection We clip and split the large tile into small images, the preprocessed dataset can be download here.
For training, you can modify parameters in configs/xx.xml, or just keep the default and:
python main.py --config='the name of config file'
e.g.
python main.py --config='the name of config file'
Please download the trained model weights here, and place them in the result directory. The evaluation and visulization are simple
# For evaluation in SParis
--config=CD_hatformer_baseline_and_BHE_FME_AFA_sparis --eval --ckpt_version bsize4_AFA --save_img
# For evaluation in SVenice
--config=CD_hatformer_baseline_and_BHE_FME_AFA_svenice --eval --ckpt_version bsize4_AFA --save_img