that is a good point. The problem is that the system that I want to build cannot run on a standard desktop or laptop computer, since it needs to have a small dimensional footprint. i am considering the intel NUC computers. If the ROS middleware adds too much overhead that performance becomes an issue, then I may have to consider other routes. I will have to miss out on the useful ROS tools.

I guess what I really am asking is: is there a minimum hardware requirement when working with high frame rate video and ROS? Maybe a potential example?

As mentioned above it completely depends on what you want to do with the data. There's very little overhead if you do nothing with the data. If you setup a large pipeline with many different sections which interact and requires copying of the data by requiring multiple write access copies to the image data at frame rate you will need a lot more resources. ROS is a tool that can do what you ask it to do. We can't estimate what resources it will take unless you can provide a description of what you actually want it to do for you. And since providing this description is hard to do accurately most people test on representative systems and try to extrapolate.

From my experience, and from the answer above, on a well designed system the majority of the computatonal time will be spent in your actual computer vision algorithm not in the interactions from any overhead of using ROS versus any other integration framework.

Personally having used a Realsense D435i, I have to say there are other considerations as well. We used an intel celeron x86 based SBC and even then the D435i was outputting so much data via USB3 that it took up one of the 4 cores available on the SBC. Mind you, this SBC was supposed to handle autonomous navigation, obstacle avoidance, object recognition, path planning, etc, and just publishing realsense data took up almost 1 core (more than 1 core if pointclouds are published). Other people I talked to also mentioned that the D435i driver was unstable due to too much data being pumped, and in the end they used it like a normal RGB camera. If you can afford it, I would highly recommend an intel NUC (i5/i7) as those guys will be able to handle the realsense better. But the nuc has limited port availability, since the camera only has a usb3 interface. It is also not known to work reliably over USB hub (power and data requirement). Realsense is good, if you can get it to work reliably.