The correct answer should be UNDEFINED, since the filter dimension should be something like NxNxNx16; i.e. it should be a 4-dimensional filter with as many channels as the input layer, that is 32x32x32x16.
Am I right?
The filter need not match the number of channels on the input image explicitly. The framework takes care of this. If you compute the number of learnable parameters of the filters, it should become clear.
We do know that the number of output channels is dependent on the filter architecture we specify.
Which framework? Are you referring to Tensorflow or similar? The question does not cite any framework; it seems to refer to general theory.
And in theory, Andrew NG explicitly explicitly note that the number of channels in a filter must match the number of channels in input layer.
His considerations directly lead to think that UNDEFINED is the expected answer.
In any case, this is just a point of view. Thank you for the answer 
Please identify exactly where he said this (lecture title and time mark). Thank you.
I ask because what he says and what he writes on the slides are sometimes different.
minute 7:14 and minute 7:39