Deeposlandia

*A nicer Pypi description*

Fixed

- From `https://github.com/Oslandia/deeposlandia/blob/master/images/...` to `https://github.com/Oslandia/deeposlandia/raw/master/images/...`

*A nice Pypi description*

Added

- `long_description_content_type` argument in `setup()` function.

Modified

- From relative image paths to absolute image paths, in `README.md`.

*Postprocessing improvement*

Added

- `--nb-tiles-per-image` as a new argument for `datagen` command.
- A progress bar for inference processes (153)

Changed

- `utils.prepare_output_folder()` returns now a dictionary of all useful output paths
- Some dependency updates (Tensorflow, opencv, pillow, keras, daiquiri)
- The preprocessing has been modified for geographic datasets: `-t`, `-v` and `-T` now
refer to raw images, the amount of preprocessed tiles being obtained by a combination
of `--nb-tiles-per-image` and these last arguments.
- The tile grid becomes optional for postprocessing (155).

Fixed

- Draws *without replacement* instead of *with replacement* in the case of preprocessing
of geographic dataset testing images (`np.random.choice` wrong parameterization). 146

Security

- `pillow` was updated to `7.1.1` (moderate severity vulnerability alert for
`pillow<6.2.2`)

Removed

- `sys.exit` statements (150)

*Packaging clean-up*

When preparing a major release, or an old release, you necessarily forget details.

Changed

- Package version 0.5 -> 0.6.1
- Long description

*Georeferenced dataset post-processing*

This release essentially copes with the georeferenced dataset, one may now post-process
the results, so as to visualize labelled masks as raster. A vectorized version of each
prediction is also available.

As another major evolution, `deeposlandia` now has a Command-Line Interface (CLI). The
available commands are `datagen`, `train`, `infer` and `postprocess` respectively for
generating preprocessed datasets, training neural networks, doing inference and
post-processing neural network outputs.

Added

- Set up a Command-Line Interface (90).
- Consider `RGBA` images and warns the user as this format is not handled by the web app
(107).
- Consider geometric treatments in a dedicated module, add vector-to-raster and
raster-to-vector transformation steps ; save postprocessed images as vector and raster
files (119).
- Postprocess aerial images so as to produce predicted rasters (118, 126, 127).
- Add missing test files for Tanzania dataset.
- Some information about GDPR in the web app (113).
- Improve unit tests dedicated to georeferenced data processing (104).

Changed

- Label folders are standardized (`labels`), in particular this folder name replaces `gt`
for `Aerial` dataset (139).
- Always use the best existing model, instead of parametrizing the access to the model
(135).
- Broken images are considered, hence not serialized onto the file system (129).
- The georeferenced aerial datasets are updated and factorized into a generic
`GeoreferencedDataset` class, the test files are updated accordingly (128).
- Deep learning model are now known as `featdet` and `semseg` instead of
`feature_detection` and `semantic_segmentation` (133).
- Update the training metric history when using a existing trained model (102).
- Move the documentation to a dedicated folder.
- Some code cleaning operations, using `black` and `flake8` (120).
- Update dependencies, especially `Tensorflow`, due to vulnerability issues.
- Fix the unit tests for Tanzania dataset population (111).
- The process quantity is an argument of `populate()` functions, in order to implement
multiprocessing (110).
- Logger syntax has been refactored (%-format) (103).

Removed

- The concept of "agregated dataset" is removed, as we consider a home-made Mapillary
dataset version. As a consequence, some input/output folder paths have been updated
(134).
- The hyperparameter optimization script (`paramoptim.py`) has been removed, `train.py`
can handle several value for each parameter (125).

*Georeferenced datasets and web application*

Some new datasets focusing on building footprint detection have been introduced in the
framework, namely Inria Aerial Image dataset and Open AI Tanzania dataset.

Some new state-of-the-art deep neural network architectures have been implemented to
enrich the existing collection, and design more sophisticated models.

Furthermore a bunch of Jupyter notebooks has been written to make the framework usage
easier, and clarify deep learning pipelines, from dataset description to model training
and inference.

And last but not least, a light Flask Web application has been developed to showcase some
deep learning predictions. Oslandia hosts this Web app at
http://data.oslandia.io/deeposlandia.