Plan

All this step must be debugged (run test) before to start another step

1. Implement 3D-FPN and data_generator
   If possible, It would be nice if the anchors are generated. The test will consist in generating the FPN. The rpn_feature_maps, maskrcnn_feature_maps will be store and analyze (graphically)

2. RPN/ROI modules
   This is the most tricky part. The natural order is:

   • RPN
   • ROI - part 1 ProposalLayer
   • ROI - part 2 DetectionTargetLayer
     After several layers of filtering the roi tensors must be check to validate this part. This part will need a NMS kernel which is not implemented in 3D (in 2D exits with Keras)

3. MaskRCNN module
   Finish the implementation. More a FC network to implement

4. Inference/detection
   Finish the job Test on Shapes3D