You may get skeptics who scoff at the new approaches and say you should learn traditional computer vision or machine learning first. I strongly believe these people are wrong.
I would expand this to include machine learning in general, with a healthy dose on deep learning architectures.
Specific tools/tech:
* Python and C++
* TensorFlow (and TensorBoard for vis)
* Keras (soon to be part of TensorFlow)
* Theano (not as popular as the previous two, but look into it)
* CUDA/CUDNN
Per the last one - if you're going to be playing in this sandbox, get used to using NVidia's tech; AMD for some reason isn't there yet - though they do have some stuff. Maybe it's a lack of marketing or something.
If you want to get your toes wet a bit - check out:
http://playground.tensorflow.org/
Finally - there's been recent research on using deep learning to create and enhance deep learning networks - so for all I know, in the future a developer role in deep learning may not even resemble an actual software development gig...
I also have to put in a plug for Geoff Hinton's Coursera course. It's not for the faint of heart. It's essentially a graduate level college course with the commensurate required time investment. It's getting a little old now but it was my introduction to neural nets and if you want to learn from a pioneer in the field it's still a good option in my opinion: https://www.coursera.org/learn/neural-networks