superpixel

A superpixel is a small, connected patch of pixels that look alike — similar color and texture — grouped together and treated as a single unit. The motivation is sheer scale: a megapixel image has millions of pixels, but only thousands of perceptually meaningful little regions. By first collapsing the image into a few hundred or thousand superpixels that respect object boundaries, you reduce the number of primitives a later algorithm must reason over by orders of magnitude, while losing almost nothing that matters.

The most widely used method is SLIC (Simple Linear Iterative Clustering, Achanta et al. 2012). It is essentially k-means run in a five-dimensional space combining color (the three Lab channels) and position (x, y). You seed cluster centers on a regular grid; each pixel joins the nearest center, but the search is restricted to a local window around each center, which keeps it fast and the superpixels roughly uniform in size. A weighting parameter trades off color similarity against spatial compactness — turn it up for blob-like regular superpixels, down for ones that hug edges tightly. Other algorithms (Felzenszwalb-Huttenlocher graph-based, watershed-based, SEEDS) make different compactness-versus-adherence trade-offs.

Superpixels were a backbone of pre-deep-learning segmentation pipelines: extract features per superpixel, classify them, and stitch the result. Their role has shrunk now that end-to-end CNNs predict per pixel directly, but they remain valuable as an efficiency layer — for fast interactive annotation (label a whole superpixel with one click), for enforcing boundary-aware smoothness, and inside graph-based methods where each superpixel becomes a node.