This PR introduces the postprocessing steps associated to semantic segmentation predictions over aerial georeferenced images.
The label inference is done on small image tiles, that must be merged once predicted labels are ready for every tile. As input, we get tiled images and the corresponding predictions. As output, we have a full labelled image directly comparable with original high-resolution image.
In more details a new module that undertake this operation is provided. It is inspired from existing inference.py module, however it focuses only on aerial image datasets (AerialImage and Tanzania, for now) and semantic segmentation models.
Additionally, a full unit test suite is provided along with the new module.
As complementary remarks, one may note that this PR introduces some duplicated code (get_image_features(.) appears twice, though it has two different signatures; then some code lines shall be merged between inference.py and postprocess.py, it may rely on issue #90). I propose to address it in a further PR.
This PR introduces the postprocessing steps associated to semantic segmentation predictions over aerial georeferenced images.
The label inference is done on small image tiles, that must be merged once predicted labels are ready for every tile. As input, we get tiled images and the corresponding predictions. As output, we have a full labelled image directly comparable with original high-resolution image.
In more details a new module that undertake this operation is provided. It is inspired from existing
inference.pymodule, however it focuses only on aerial image datasets (AerialImageandTanzania, for now) and semantic segmentation models.Additionally, a full unit test suite is provided along with the new module.
As complementary remarks, one may note that this PR introduces some duplicated code (
get_image_features(.)appears twice, though it has two different signatures; then some code lines shall be merged betweeninference.pyandpostprocess.py, it may rely on issue #90). I propose to address it in a further PR.