CNN Convolution Explorer

2D Convolution

A filter (kernel) slides over the input image with a given stride $S$ and padding $P$. At each position it computes a dot product with the overlapping patch, writing one value into the feature map. Use the controls below to explore how each parameter changes the result.

Input size
7×7
with padding: 7×7
Filter size
3×3
receptive field
Output size
5×5
feature map
Position
row, col (0-indexed)
Image
Filter K 3
Stride S 1
Padding P 0
Pad value
Speed 4
step /
Input Feature Map 7×7
conv2d
Output Feature Map 5×5
Filter / Kernel (click cells to cycle values)
Presets
Current dot product computation
Press Play or Step to begin.
Tip: Start with Padding = 0, Stride = 1, Filter = 3. Then increase padding to see how the "virtual border" expands the input — with zero padding (black) the border is 0, with one padding (white) it is 1. Increase the stride to watch the kernel jump further, producing a smaller feature map.