A hands-on route from sliding weighted neighborhoods to output shapes, stride, padding, filter count, and pooling.
A cricket speed gun captures a short burst of rapid readings for each ball. A 3-reading weighted window slides over that burst, and the same sliding-window idea then extends to 2D neighborhoods.
Each delivery has five quick readings. A 3-reading window slides across them, giving the newest reading in the window the largest weight.
The same original image can be transformed in different ways depending on the filter weights. In CNNs, these useful filters are learned from data instead of hand-picked.
A filter is a small pattern detector or image operator that is reused across the whole image.
A small practice-net scene with edges, texture, and smooth regions. The three filters below all start from this same image.
Reduces small noisy changes and softens detail. Useful before later processing when tiny variations should not dominate.
Highlights where brightness changes sharply. Useful for locating outlines such as the ball, bat, crease, or object boundaries.
Boosts local contrast so details look crisper. Useful when boundaries and fine texture need to stand out more strongly.
Walk through the lecture sequence: normal filter shrinkage, stride, padding and floor, depth from K filters, then pooling.
With stride 1 and no padding, the filter cannot start where it would fall outside the input.
A 3 x 3 filter loses one border cell on every side, so 6 becomes 4.
Stride does not change the filter. It changes how many legal starts you actually visit.
Stride 2 roughly halves the number of starts. Floor keeps only full filter placements.
Padding gives boundary cells enough artificial neighbors, then stride and floor decide the final count.
Padding adds P cells on every side, so the width gains 2P before filtering.
Partial border placement skipped by floor.
A filter spans the full input depth, slides across width and height, and produces one 2D feature map. K filters produce depth K.
Each filter has F x F x D1 weights.
Pooling works per feature map. Max pooling keeps the strongest local response; average pooling keeps the local mean.
Pooling has 0 learned parameters and keeps depth K.