Note that in your first statement, you are reacting to this:

"The compilers, for the most part, are smarter than people at optimizing code."

And then you say:
"But a competent assembler specialist (someone in the same league, skillwise, with the guys that write the compiler)[...]"

Put it together: there are a whole lot of programmers, and there are a handful of skilled assembler specialists. So if you take a random coder, and force him to code in assembly, your chances to gain performance is damn small.

Although you are trivially right about the skilled people, the statement you answered was not about them, that's why it says "for the most part".

There is a common misconception which says "if it is too slow, code it in assembly", which should sound like "if it is too slow, learn CPU tweaking for two decades, and then code it in assembly" :) In the old 8bit or 8086 days, without SSE+FPU, pipelines, cache coherence and branch predictors, assembly coding was a pretty straightforward business, the coder could do nice things by keeping things in registers, and knowing basic idioms as XOR AX AX, and in the same time compilers were tuned to create the binary fast, there were little time for optimization. Basic knowledge was enough to create faster code than compiler.

Kinect is uneffective in daylight. They have to use a custom built kinect variant in order to get the infra camera to sense the projected dots in sunshine. The sun is quite a rival as source of infra light :) It also have hard time detecting shiny surfaces. The dangerous objects and situations make a fine list for the users :)

Newtonian physics say: F=ma. Not F=ma+God's will. To put this way, anybody who think newtonian physics is right and working, implicitely denies any wonders.