I need to build a detector for a new class, and then run it using the DeepStream and the TensorRT framework.
So - I have annotated images, and need model that return BB and probability for the item i wish to detect. To train the model, I usually to work with Keras, or if there is no other choice - tensorflow.
Not sure what is the best option for running something on the DeepStream/Tensorrt. Naturally, I would use SSD on Keras. But it seems non trivial to train it and then convert to the the tensorrt engine.
Should I use a simpler, optionally slower network? or is there some “cookbook” to do that ?