Computer vision / real pixels, local filters

Convolution Lab

Run one editable 3×3 CNN-style cross-correlation over real photos. Switch between depthwise RGB and a true B/W luminance image, then inspect the exact pixels behind any output response.

Choose a real image

Bundled photos work offline. Uploads are decoded and processed only in this browser.

JPEG · PNG · WebP · max 12 MB

Pixel treatment

RGB applies this same spatial kernel independently to red, green, and blue, then visualizes the three responses together. B/W first computes Y = 0.2126R + 0.7152G + 0.0722B and produces one signed response map. CNN libraries conventionally call this cross-correlation operation convolution, even though the kernel is not flipped here.

The kernel controls

A single filter is deliberately small enough to edit by hand.

Kernel recipe

Stride1

Move the kernel 1 working pixel between outputs.

Zero padding1

Padding lets the filter visit a photo's edge pixels.

ReLU, then max-pool luminanceFirst clamp negative responses to 0. Then keep the largest remaining activation in each 2×2 region.

Photo field station

Click a feature-map pixel or jump to a real extreme. The enlarged lens below follows that exact output position.

Sampling City geometry

Building a compact working raster…

Images are downsampled once to a maximum 256px edge so the selected receptive field remains interactive. Uploaded files never leave this browser.