Image matting

Count the unknowns. Each pixel gives you three measured numbers, red, green and blue. To solve the compositing equation you need seven: three for the foreground colour, three for the background colour, and one for alpha. Three equations, seven unknowns, so the problem is under-constrained and has no single correct answer from the pixel alone.

Segmentation avoids this entirely by refusing to answer it. It picks a side per pixel and moves on. Matting has to lean on priors, neighbouring pixels and learned statistics about how real edges behave, which is why matting models are fussier, slower and much more sensitive to input quality.

A trimap is a three-value hint map: definitely foreground, definitely background, and unknown. It shrinks the problem to a narrow band along the edges, where the solver only has to work out alpha inside the unknown strip. Classical matting depended on a good trimap, and drawing one by hand is tedious.

Trimap-free models generate that structure internally, usually by running a coarse segmentation first and treating the boundary region as the unknown band. They are far more convenient and handle most photos well. On genuinely hard images, wispy hair against clutter, a hand-drawn trimap still beats an automatic one.

Anything where the edge is not an edge but a transition:

• Hair and fur, where individual strands are thinner than a pixel
• Smoke, steam and flames, which are semi-transparent throughout
• Glass, acrylic and clear packaging, where the background is genuinely visible through the subject
• Veils, lace, mesh and knitwear, with holes too small to resolve
• Motion-blurred limbs, where the edge is smeared over many pixels

For a hard-edged product on a plain backdrop, a clean mask is often enough. The moment a subject has any of the above, a mask alone will look cut out with scissors.