Codes for the HIDA/IDSI international archaeology AI challenge where we won first place among the participants. The task is multiclass semantic segmentation of ancient agricultural terrace/wall in the Negev desert. We use UNet/DeepLabV3+ with EfficientNet backbone and get an overal IOU (foreground only) of 0.31.
Orthophoto | Prediction | GroundTruth |
- The training data are 500 512*512 image patches, each patch containing 9 feature images (Orthophoto, Aspect, DTM, Flow_Accum, Flow_Direction, Prof_curv, Slope, Tang_curv, Topo_Wetness).
- The testing data (for final score) are 200 unseen image patches.
- architecture: UNet/DeepLabV3+, we merge the output probability map of both models to calculate the final prediction
- backbone: efficientnet-b5 (pretrained on ImageNet)
Codes for the models are based on segmentation-models-pytorch.
Cross entropy loss + DICE loss, AdamW optimizer, cosine loss decay
- train:
python main_train.py
orbash train.sh
- inference:
python main_inference.py
orbash inference.sh
(python main_inference_2model.py
to do inference with merged models) - evaluate:
python main_test.py
@inproceedings{wang2022deep,
title={Deep Semantic Model Fusion for Ancient Agricultural Terrace Detection},
author={Wang, Yi and Liu, Chenying and Tiwari, Arti and Silver, Micha and Karnieli, Arnon and Zhu, Xiao Xiang and Albrecht, Conrad M},
booktitle={2022 IEEE International Conference on Big Data (Big Data)},
pages={4888--4892},
year={2022},
organization={IEEE}
}