Deploy trained models generated by Deep Learning frameworks
The classification and regression filters and applications provided by OTB are only able to use models generated by OTB's training tools (which use OpenCV, Shark or custom OTB models).
With the current DL hype, the models are developed and trained using frameworks like Keras leveraging Tensorflow for instance. OpenCV provides the ability to import Caffe and Tensorflow models.
The same functionalty should be available in OTB in order to deploy remote sensing DL at scale.
It seems that OpenCV has taken the option of not depending on Tensorflow or Caffe but just on protobuffer. This means that they read the models from the protobuffer format but reimplement all the needed layers. This may limit the kinds of models that can be deployed or need to develop custom layers (https://docs.opencv.org/trunk/dc/db1/tutorial_dnn_custom_layers.html).
If OTB moves to OpenCV 3.4 all this comes for free.
The other option is to depend on Tensorflow (just the static C API) but being able to import any model trained outside of OTB.