Run inference on binary (2 classes only) Deeplab model cropping large images in a sliding window approach.
You will need the following things properly installed on your computer:
- Git
- Python >= 3.6
- pipenv or docker
git clone git@github.com:ttill/deeplab-inference.gitcd deeplab-inference
pipenv install
docker build -t deeplab-inference .
Usage: app.py [OPTIONS]
Run inference on binary (2 classes only) Deeplab model stored in frozen
graph on image. Image has to be in RGB format. It will be cropped in a
sliding window approach (see option `crop_size`).
Options:
--frozen-graph FILE Path to frozen graph of the model to use in
protobuf format. [required]
--input PATH Path to input image to infer on. In case a
directory is passed inference will be run on
every contained. [required]
--crop-size INTEGER In case input is larger in terms of width or
height, it will be cropped into patches of
`crop_size` with overlaps of 50 pixels will
be passed into the model. Predicitons will
be stitched together considering the maximum
value in the overlapping regions.
--segmentation PATH Where to put a PNG image of the segmentation
map. Filename or folder (uses basename of
input).
--probability PATH Where to put a PNG pseudocolor image of the
probability map. Filename or folder (uses
basename of input).
--visualize-progress / --no-visualize-progress
Show figure containing input, probability
map (pseudocolors), segmentation overlay for
every cropped patch (halts inference).
--visualize-result / --no-visualize-result
Show figure containing input, probability
map (pseudocolors), segmentation overlay for
every input image (halts inference).
--ground-truth PATH Path to image (filename or folder where png
with same basename as input is located) with
ground truth as 8-bit grayscale png (class
0: `0`, class 1: `255`). When supplied
multiple metrics will be calculated
(Accuracy, MIoU, Matthews Correlation
Coefficient, …)
--split-file FILE Use this to filter input folder. Path to
text file with one basename per line (pascal
voc format).
--difference PATH Where to put a PNG image containing a
difference map: Black: True negative (class
0), White: True positive (class 1), Red:
False negative, Blue: False positive.
Filename or folder (uses basename of input).
--install-completion Install completion for the current shell.
--show-completion Show completion for the current shell, to
copy it or customize the installation.
--help Show this message and exit.
pipenv run python app.py --frozen-graph path/to/frozen_inference_graph.pb --input path/to/images/ --segmentation output/folder/
docker run -it -v /path/to/data/:/data/ deeplab-inference bashpython app.py --frozen-graph /data/frozen_inference_graph.pb --input /data/images/ --segmentation /data/segmentation/
This project is licensed under the MIT License.