The only one risky thing is that in computer vision right now things change each 6 months. So, one will never know which techniques are still state of the art, and which were outclassed by some ConvNets 4 years ago.
- considered obsolete (e.g. complicated and nasty, and still outclassed by NNs),
- building blocks of current techniques,
- provided inspiration for current techniques,
- are still state of the art.
Mostly what's changed is that what used to be heuristics and hand-crafted features is now getting replaced by proper learned models. Also some generative models are getting replaced discriminative models since deep learning does a very good job in creating those.
SLAM using SfM (from monocular vision) using Deep Nets would be huge. Would reduce the specifications for LIDAR in SDC's for example.
Even if you don't go into computer vision - these have a wide range of applications, for example in sound or signal processing which are closely related.