Jane Hung

We should remove small objects at the edge (could be by side length, % of image side, or remove any object at the edge) https://github.com/broadinstitute/keras-rcnn/blob/ad6f05fa2a4510a310f37b8a73548614dc2f8df8/keras_rcnn/preprocessing/_object_detection.py#L109

The image pixels are limited to range 0 to 1, which is an issue if you want to center around 0 https://github.com/broadinstitute/keras-rcnn/blob/4024b14711338423fba30acd7c741ae63e812aac/keras_rcnn/preprocessing/_object_detection.py#L342

Using Keras' validation loss (i.e. setting validation_data parameter in model.fit_generator) could lead to an Unimplemented error, so don't use it

NaN loss when object has category not in dictionary

4

If the dictionary of categories does not contain a category that an object has, the loss becomes NaN

Mean average precision (mAP) metric

2

Resolves #6

We need a central location where parameters are stored. This can be done through a config dictionary, e.g. config = {max_proposals: 300}

Dictionary schema for datasets

3

Make a standard dictionary schema for all datasets and make sure existing ones (malaria, pascal) adhere.

Add inside and outside bounding box weights

2

As in https://github.com/rbgirshick/py-faster-rcnn/blob/master/lib/fast_rcnn/config.py#L120 and used in https://github.com/rbgirshick/py-faster-rcnn/blob/master/lib/rpn/proposal_target_layer.py#L64 and regression losses.