Lightweight Mapping Model

Overview

---

This is a demo of OpenEarthMap lightweight models searched with SparseMask and FasterSeg neural architecture search methods. The models were automatically searched and pretrained on the OpenEarthMap dataset (using only the training and validation sets).

OpenEarthMap dataset

---

OpenEarthMap is a benchmark dataset for global high-resolution land cover mapping. OpenEarthMap consists of 5000 aerial and satellite images with manually annotated 8-class land cover labels and 2.2 million segments at a 0.25-0.5m ground sampling distance, covering 97 regions from 44 countries across 6 continents. OpenEarthMap fosters research, including but not limited to semantic segmentation and domain adaptation. The project website is https://open-earth-map.org/

@inproceedings{xia_2023_openearthmap,
    title = {OpenEarthMap: A Benchmark Dataset for Global High-Resolution Land Cover Mapping},
    author = {Junshi Xia and Naoto Yokoya and Bruno Adriano and Clifford Broni-Bediako},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month = {January},
    year = {2023}
}

Lightweight model

---

The lightweight models searched and pretrained on the OpenEarthMap dataset can be downloaded as following:

Method	Searched architecture	Pretrained weights	#Params	FLOP
SpareMask	mask_thres_0.001.npy	checkpoint_63750.pth.tar	2.96MB	10.45GB
FasterSeg	arch_1.pt	weights1.pt	3.47MB	15.43GB

Usage

---

• SparseMask model: download the architecture mask and the pretrained weights and put them into folder models/SparseMask/.
  Start the evaluation demo as:

python eval_oem_lightweight.py \
    --model "sparsemask" \
    --arch "models/SparseMask/mask_thres_0.001.npy" \
    --pretrained_weights "models/SparseMask/checkpoint_63750.pth.tar" \
    --save_image --save_dir "results" 

       Or use the Jupyter notebook: sparsemask_demo.ipynb.

• FasterSeg model: download the architecture structure and the pretrained weights and put them into folder models/FasterSeg/.
  Start the evaluation demo as:

python eval_oem_lightweight.py \
    --model "fasterseg" \
    --arch "models/FasterSeg/arch_1.pt" \
    --pretrained_weights "models/FasterSeg/weights1.pt" \
    --save_image --save_dir "results" 

       Or use the Jupyter notebook fasterseg_demo.ipynb.

Example of predictions

---

• SparseMask model
  
  
• FasterSeg model
  
  

Acknowledgement

---

Automated neural architecture search method code from

• SparseMask
• FasterSeg