Automatic detection and semantic image segmentation with deep learning
This project aims at showcasing some Deep Learning use cases in terms of image analysis, especially regarding semantic segmentation.
The project contains the following folders:
- deeposlandia contains the main Python modules to train and test convolutional neural networks
- docs contains some markdown files for documentation purpose
- examples contains some Jupyter notebooks that aim at describing data and building basic neural networks
- images contains some example images to illustrate the Mapillary dataset as well as some preprocessing analysis results
pytestis used to launch several tests from this folder.
Additionally, running the code may generate extra subdirectories in the chosen data repository.
The code has been run with Python 3. The dependencies are specified in
setup.py file, and additional dependencies for developing purpose are listed
$ git clone https://github.com/Oslandia/deeposlandia $ cd deeposlandia $ virtualenv -p /usr/bin/python3 venv $ source venv/bin/activate (venv)$ pip install -r requirements-dev.txt
As a particular case, GDAL is not included into the
Ubuntu distributions, the following operations are needed to install this
sudo apt-get install libgdal-dev sudo apt-get install python3-gdal
GDAL version can be verified by:
After that, a simple
pip install GDAL may be sufficient, however considering
our own experience it is not the case on Ubuntu. One has to retrieve a
for Python that corresponds to the
GDAL of system:
pip install --global-option=build_ext --global-option="-I/usr/include/gdal" GDAL==`gdal-config --version` python3 -c "import osgeo;print(osgeo.__version__)"
For other OS, please visit the
GDAL installation documentation.
Running the code
A command-line interface is proposed with 4 available actions (
postprocess), callable as follows:
deepo [command] --options
Some files document the command use:
- Preprocessed dataset generation
- Train a model
- Infer labels
- Postprocess results for geographic datasets
- Run your own web app instance
In this project we use a set of images provided by Mapillary, in order to investigate on the presence of some typical street-scene objects (vehicles, roads, pedestrians...). Mapillary released this dataset on July 2017, it is available on its website and may be downloaded freely for a research purpose.
As inputs, Mapillary provides a bunch of street scene images of various sizes
images repository, and the same images after filtering process in
There are 18000 images in the training set, 2000 images in the validation set, and 5000 images in the testing set. The testing set is proposed only for a model test purpose, it does not contain filtered versions of images. The raw dataset contains 66 labels, splitted into 13 categories. The following figure depicts a prediction result over the 13-labelled dataset version.
In the Aerial image dataset,
there are only 2 labels, i.e.
background and consequently the
model aims at answering one single question for each image pixel: does this
pixel belongs to a building?
The dataset contains 360 images, one half for training one half for
testing. Each of these images are 5000*5000
tif images. Amongst the 180
training images, we assigned 15 training images to validation. One example of
this image from this dataset is depicted below.
Open AI Tanzania
This dataset comes from the Tanzania challenge, that took place at the autumn 2018. The dataset contains 13 labelled images (2 of them were assigned to validation in this project), and 9 additional images for testing purpose. The image resolution is very high (6~8 cm per pixel), that allowing a fine data preprocessing step.
In such a dataset, one tries to automatically detect building footprints by distinguishing complete buildings, incomplete buildings and foudations.
To complete the project, and make the test easier, a randomly-generated shape model is also available. In this dataset, some simple coloured geometric shapes are inserted into each picture, on a total random mode. There can be one rectangle, one circle and/or one triangle per image, or neither of them. Their location into each image is randomly generated (they just can't be too close to image borders). The shape and background colors are randomly generated as well.
How to add a new dataset?
If you want to contribute to the repo by adding a new dataset, please consult the following instructions.
This project implies non-commercial use of datasets, anyway we can work with the dataset emitters to get commercial licences if it fits your demand. May you be interested in any pre-trained models, please contact us at email@example.com!
The program license is described in LICENSE.md.
Oslandia, April 2018
Release history Release notifications | RSS feed
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size||File type||Python version||Upload date||Hashes|
|Filename, size deeposlandia-0.6.3.post1-py3-none-any.whl (57.9 kB)||File type Wheel||Python version py3||Upload date||Hashes View|
|Filename, size deeposlandia-0.6.3.post1.tar.gz (46.7 kB)||File type Source||Python version None||Upload date||Hashes View|
Hashes for deeposlandia-0.6.3.post1-py3-none-any.whl