ArtifiVe-Potsdam

tooling and demo for ArtifiVe-Potsdam (Artificial Vehicles) dataset http://rs.ipb.uni-bonn.de/data/artifive-potsdam/

Please check the notebook to see how the code can be used to create a pytorch dataset, to prepare the data with transformations, and some visual samples of the dataset content.

real (Potsdam) samples

based on ISPRS 2D Semantic Labeling Contest - Potsdam:

artificial samples

based on blueprints:

Installation

Download the content of the repository and run the following to install the requirments:

pip install -r requirements.txt

This installs all requirements with pip, except of libjpeg-turbo, which can easily be installed with:

conda install libjpeg-turbo -c conda-forge

Dataset

• the annotations are provided as a single json file for each sub dataset called annotations_*.json
• the content is a list of json dictionary for each sample (image + annotations) in the dataset
• labels are stored in form of the class strings
• objects are stored in form of their bounding polygons and in WKT format, which can for example be converted using shapely's shapely.geometry.shape()
• a is_difficult flag is also contained, but is allways false

dataset	training images	test images	training objects	test objects
fullsized	24	14	6019	3833
patched 600x600 px	2400	1400	6978	4489
patched 600x600 px+ 200 px overlap	5400	3150	15379	9793
artificial	1000		10000

Training + Test split

The split is based on the original ISPRS 2D Semantic Labeling Contest split. The green colored tiles are the test section.

Requirements

• numpy https://github.com/numpy/numpy
• pytorch https://github.com/pytorch/pytorch
• torchvision https://github.com/pytorch/vision
• pillow https://github.com/python-pillow/Pillow
• shapely https://github.com/Toblerity/Shapely
• matplotlib https://matplotlib.org/
• optional: libjpeg-turbo https://github.com/libjpeg-turbo/libjpeg-turbo/ + https://github.com/lilohuang/PyTurboJPEG

Benchmark details

Training

• we train on patched/600x600_overlap200/training
• we split it into 70% for training and 30 % for validation, therefore the provided sample numbers in the paper are smaller than the actual number of images; also therfore your performance may vary
• we remove objects whose min and max sizes are outside of 20 and 200 px and require one side to be larger than 40 px
• we remove the empty images which further reduces the number of samples (remove_empty sample_filter)

Testing

• we test on patched/600x600/test dataset
• we use pycocotools to evaluate
• the baselines report the AP with IOU threshold 0.5