UAV-segmentation scripts

Collection of scripts for preparation of datasets for semantic segmentation of UAV images

Those scripts were used to prepare and unify multiple datasets used to train a semantic segmetation model used in an Automatic Safe Emergency Landing application. It converts all of the supported datasets to a cityscape type structure. This is useful as most semantic segmentation repos out there support Cityscape. On top of this, this repo supports re-labelling. This means you can group and discard annotations. This is particularly useful when combining several datasets together.

User Guide

1. Download one/several of the datasets supported.
2. Update the configuration file:
   • DATASETS_PATHS: this should contain the root directory to the datasets.
   • size_out: all the labels/images will be resized to the value you enter here.
   • use_default_split: If True, the data splits given by the dataset authors will be used. Otherwise (or if the authors did not specify sets), a split will be generated (70/15/15). You can change the proportion of the data splits in dataset_prep.py
   • use_train_ids: if True, the trainIds will be used instead of the labelIds. For more details on the difference between the two. Please refer to labels.py
3. (optional) Create virtual environment
4. pip3 install -r requirements.txt
5. python3 setup.py install
6. Run python3 scripts/run.py --dataset [insert dataset_name]
7. Feel free raise an Issue, star, or do PRs :)

Supported Dataset

Dataset	Number of images	Number of classes	Description	Perspective
Cityscape (2D fine)*	5000 (50 seq)	30	High-quality autonomous driving segmentation (not UAV)	Ground Level
Aerospace	3269	11	images captured from commercial drone (5 to 50 meters altitude)	Mixed (forward/down)
TU-GRAZ landing	592	20	Segmentation for Urban environment for safe landing	down
UAVid	420 (42 seq)	8	Segmentation for Urban environment	forward

*Cityscape still being implemented. However, you can just refer to this repo in the meanwhile.

Common Labels across datasets

Dataset \ Labels	road	person	vehicle	building	vegetation	background
Cityscape (not UAV)	✔️	✔️ (person/rider)	✔️ (bus/truck/car)	✔️	✔️	✔️
Aerospace	✔️	✔️	✔️	✔️	✔️	✔️
TU-GRAZ landing	✔️ (gravel)	✔️	✔️	✔️ (door/window/roof/wall)	✔️(tree/gras)	✔️
UAVid	✔️	✔️	✔️ (static/dynamic)	✔️	✔️ (tree/low veg.)	✔️

Credits

Some of the code (particularly label.py) was based on this repo.