At the very first phase when you give empty domain it is non-conditional failure - it is the same as saying: "I have a function $f(x) = x^2 + 1$, but $x \in {\emptyset}$, find me feasible solutions, which are none.
In the phase when arcs are getting cut you check if the values are consistent between two variables - here arc is either in solution or is inconsistent.
You have the same function as above and points: $x \in \{0, 1\}, y \in \{1, 5\}$. This makes one consistent pair, the second is cut, it will not be in the solution. At this point if there were more constraints you would check the consistency of related variables (You might call it backtrack, because another variable that is related takes part, but here you might find that there are no arc, and it is perfectly fine here). But if $x \in \{23, 45\}$ then you cannot find any good pair, but now it turns out that domains of $x$ is $\emptyset$, so something was inconsistent - with checked constraints you have no feassible points, so it is failure.
The sildes you have provided show that AC-3 is detecting wrong candidate for solution faster than forward checking. The outer scheme is backtracking solver, which at every step assumes one of possible values, assigns it and continues guessing values until it finds solution or contradiction (in such case it goes back to assumption made and changes it to different possible value). Slides show that propagating constraints is superior to propagating values, because it allow faster checking of error.
AC-3 in this context is used as black-box function to determine failure, but the pure algorithm itself does not backtrack - it operates on one given instance - candidate for solution, and it's output serves as indicator to backtracking.