So far, we've seen the pooling layer operations, max and average pooling, with a two-dimensional matrix. But, as we saw, RGB images and color images are three-dimensional matrices, with three channels, so the value of the third dimension is three:

Similar to the convolution, we'll also take three filters, and for each of the filters we'll do the operation we saw in the previous section independently. The filter will handle the pooling layer operation with the channel. It will, of course, produce three two-dimensional matrices, which will be the product of the three independent operations.

Now, if you recall from the convolution, because those matrices were no longer presenting colors, we added those matrices together, and therefore had only one matrix in the end. Well, this final step won't be done in the pooling layers. The pooling layers aren't adding those matrices together, but leaving them as they are.

So, the third dimension was 3, and with max pooling, for example, the output will also be 3. The third dimension is left untouched. If it were a convolution layer, we'd still have three filters, but the third dimension would be cut to 1.