An interesting comparison (maybe) is the LIDAR unit that was developed for the Neato XV series robot vacuum cleaners.
When it was first announced, and the research paper was published showing how it was built and functioned (essentially a 2D spinning parallax distance measuring system using a laser and linear photosensor array), it was thought that the system would be very cheap to make, and would eventually sell to hobbyists for around $25.00.
Unfortunately - that never happened.
I don't know what it ultimately cost Neato to make; it was likely inexpensive in quantity, but it was never sold to hobbyists or robotics researchers as a standalone product. Instead, if you wanted one, you had to either purchase and cannibalize one of the robots, or otherwise obtain one as a replacement or "pulled" component. Even then you would spend much more than $25.00 to get one.
Eventually they started to show up on Ebay; at one point I was able to purchase the units for around $35.00 at the cheapest; these were "used pulls", some needed new motors, pulleys, and/or drive belts (all parts easily found as well) - so that pushed the price up a bit.
Lately, the prices on Ebay have settled to around $75-90.00 for refurbished pulls, or replacement new units.
Interestingly, similar units have also appeared on "Chinese clone" vacuum robots, and you can purchase the LIDAR units they use separately on sites like Alibaba, etc. But even there they don't sell for anywhere near $25.00 each.
I am not sure where the discrepancy for this device lies; maybe the quantity just isn't there, and honestly $100.00 for a 2D LIDAR (granted, indoor only and very limited range) is fairly inexpensive, but even so I don't see many people using them on their hobbyist devices. Most research robotics continue to use the more expensive SICK and Hokuyo 2D LIDAR (which are time-of-flight based).
I think it is more likely that if LIDAR is used, it will be part of a large suite of sensors, and likely 2D only, unless the price of flash depth LIDAR devices plummet.
I'm not an expert, but I personally think that computer vision based approaches can work well, especially combined with other sensor systems (LIDAR, RADAR, ultrasonics, etc). NVidia has shown that vision-based approaches can be made to work, and others have done similar work that show the approach to be viable. I'm a bit biased on this because I used a simplified version of NVidia's approach to pilot a simulated vehicle in an online MOOC I was a part of a few years back (I had to simplify the CNN to make the model fit the memory I had in my GPU at the time; I had tried to use a batch training method but wasn't able to get it to work properly before my deadline - even so, the simplified model trained and generalized quite well with the data I gathered).