where is the starting point

Hi, @akshaylamba!

keras-rcnn is still a work-in-progress. I wouldn’t recommend experimenting with the package for a few weeks unless you’d like to contribute! If you’re interested in contributing, I recommend starting with the aforementioned malaria dataset and the ObjectDetection generator:

import keras_rcnn.datasets.malaria

training, test = keras_rcnn.datasets.malaria.load_data()

generator = keras_rcnn.preprocessing.ObjectDetectionGenerator()

generator = generator.flow(training)

validation_data = keras_rcnn.preprocessing.ObjectDetectionGenerator()

validation_data = generator.flow(test)

The ObjectDetection generator will provide the anchors needed for training, e.g.

image, (labels, bounding_boxes) = generator.next()

They look like this:

>>> image.shape
(1, 800, 600, 3)

>>> bounding_boxes.shape
(1, 50, 38, 18)

>>> labels.shape
(1, 50, 38, 72)

After you’ve generated a training sample, it’s more complicated. We’re in the process of refactoring our region proposal network (see the losses pull request). But in the next few days you should be able to pipe the output of the ObjectDetectionGenerator directly to the RegionProposalNetwork model.

https://github.com/broadinstitute/keras-rcnn/issues/7 provides some details on a possible implementation.

If you have opinions about the design or architecture, please let us know! We’re actively looking for feedback. We aim to make this a package that’s made by and for the computer vision and Keras communities!

broadinstitute / keras-rcnn

where is the starting point #36