We are evaluating Deepstream to see if it could fit with our use cases and we are having a number of questions that we can’t answer with the documentation.
Is it possible to create a flow with only a classification model using the Python Bindings? Is it necessary to modify the code and recompile Deepstream and modify the Python Bindings for this? If possible, what is the best way, using nvinfer or nvinferserver?
In the nvinferserver documentation it is stated that only multi-class classification, detection and segmentation models can be used, is this correct, is this a limitation of nvinferserver or is it a limitation of Triton?
Is it possible to use a custom model with the Python Bindings such as keypoints, line detection or depth estimation? Is there any documentation or can you give me some guidance on what steps to follow?
Is it possible to incorporate post-processing logics on the predictions using the Python Bindings such as, for example, using the positions of the different detections of a vehicle to calculate its velocity or to make a projection of its position?
Is it in the deepstream or TAO toolkit roadmap to incorporate new models such as, for example, depth detection? Is there a public roadmap to consult?
I would appreciate any full or partial answer.