In the sample models, there is the person-reidentification-retail-0076 - which sounds very useful. I'm learning, so please forgive my ignorance.
For the record, I'm using an NCS2.
Do such models have to be trained on people it should recognize, or is the model inferring an object then maintaining that to the next inference, and providing the seen before?
How do I create such a model for something other than people?
I'm trying to detect, cassify, and capture vehicles, and sometimes they move across the image quickly due to speed, so to avoid as many duplicate captures as possible, I need to track them.
This type of network seems like it would suit my needs, but I don't understand how it works or how to construct it.
I was wondering if I could retrain this network for vehicles.
So in digging - and I'm new at this - could I take the prototxt file - and modify the input for the network to be, instead of 96 x 48, make it 48 x 96 - then, rerun the optimizer and produce a new bin/xml network?
In this way, would I then have a(n) reidentificaiton network that favors objects that are rectangular and horizontal?