# File structure
VMA
├──data
│ ├── nyc
│ │ ├── cropped_tiff
│ │ ├── labels
│ │ ├── data_split.json
Run
# script to prepare tiff images
sh ./tools/icurb/get_data_new.bash
# script to prepare labels
# install gdown
pip install gdown
# download and unzip the label from Google Drive
sh .tools/icurb/get_label.bash
We directly download the labels and the raw aerial image data from NYC database. Then we conduct a set of processings to obtain tiff images that we need. It may take some time for download and processing.
In case the script fails, you can download and unzip the data manually here. Then use the script in ./tools/icurb to process them.
./tools/icurb/data_split.json defines how the dataset is split into pretrain/train/valid/test. They are randomly split. It is recommended to use our provided data splitting file.
We provide the remapped iCurb backbone of our implementations. You can download it in Baidu/Google and saved in ./ckpts.
We provide some sd dataset for train and eval, you can download them in Baidu/Google and place them as shown below.
# File structure
VMA
├── data
| ├── sd_data
| | ├──line
| | | ├──origin_data
| | | | ├──image_data
| | | | ├──trajectory_data
| | | | ├──line_6k_data.json
| | ├──box
| | ├──freespace
Use the script in ./tools/custom to crop the orginal 6k data. The cropped data and annotation file will be generated in sd_data/line/cropped_data. Then you can train these data with the configs