Efficiently consolidate matches by class ID in numpy

Question

I'm working on a machine learning task where I'm correlating a set of (Ni) instances with a set of (Nr) annotated regions, that correspond to particular (Nc) classes.

As an analogy, let's imagine we have a satellite picture with lots of animals in. We already know where all the animals are, but somebody's drawn boxes/circles/polygons/whatever around groups of them - to tag them as particular species (horse, cow, chicken, etc).

I already have:

• A matches (Ni, Nr) matrix which describes whether each instance falls within each annotated region
• A class_ids (Nr,) array describing the class ID of each region (0 to Nc-1)

I'm looking for an efficient way to get to an (Ni, Nc) result matrix which is 1 where any annotated region for class c captured instance i - and 0 otherwise.

It feels like there should be some magic indexing/assignment operator(s) that would help with this but I'm just not seeing it.

For example the following doesn't quite work:

# Example inputs
Nc = 6  # (Nr = 3, Ni = 5 below)
class_ids = np.array([2, 5, 5])
matches = np.array([
    [0, 0, 0],
    [1, 0, 0],
    [0, 1, 0],
    [0, 1, 1],
    [1, 1, 0],
])

# Non-working method:
result = np.zeros((Ni, Nc))
result[:, class_ids] = matches

# Result:
# Sets `i, c` to 1 only where the *last* annotation of type `c` in the list
# matched that instance.
array([
    [0., 0., 0., 0., 0., 0.],
    [0., 0., 1., 0., 0., 0.],
    [0., 0., 0., 0., 0., 0.],  # Final should be 1
    [0., 0., 0., 0., 0., 1.],
    [0., 0., 1., 0., 0., 0.],  # Final should be 1
])

I understand of course there are for loop-y ways of doing this by starting with result = np.zeros((Ni, Nc)) and then iteratively patching it from the matches, but felt like there should be some better, direct/vectorized way. Maybe there is some option to do a partitioned max on matches and then the assignment? Hope some numpy wizard can see something I'm missing