Training the HoliCity V1 through MaskRCNN (Detectron2).

See INSTALL.md.

Downloading the HoliCityV1 dataset from the homepage holicity.io, which includes split-v1, image, plane . Unzip into the folder dataset/ and reorganized as follows: (The clean-filelist.txt already existed in the folder dataset/. )

dataset/
    image/
        2008-07/
        2008-09/
        ...
    plane/
        2008-07/
        2008-09/
        ...
    split/
        v1/
            clean-filelist.txt
            filelist.txt
            train-middlesplit.txt
            test-middlesplit.txt
            valid-middlesplit.txt

You can download our reference pre-trained models from Google Drive. Those models were trained with HoliCity/init.py for 100k iterations.

The default batch size assumes your have a graphics card with 8GB video memory, e.g., GTX 1080Ti or RTX 2080Ti. You may reduce the batch size if you have less video memory.

CUDA_VISIBLE_DEIVCES=0 python main.py -s train -m HoliCityV1

It will build the train (HoliCityV1_train_coco_format.json) and valid (HoliCityV1_valid_coco_format.json) json file in the folder data/HoliCityV1_v1/ first. (It will cost about 1.5 and 0.5 hours respectively.)

To test the pretrained MaskRCNN above on your own images, you need change the HoliCity/init.py

self.ckpt = /the/checkpointfile/you/trained/

def predict(self):
    self.img_dirs_list = [the paths list of your own images]

and execute

CUDA_VISIBLE_DEIVCES=0 python main.py -s predict -m HoliCityV1

• Xili Dai
• Yichao Zhou

This project is licensed under the MIT License - see the LICENSE.md file for details

• Yichao Zhou