Matting evaluation shows region and normalization choices can shift MSE more than sixfold
A small matting evaluation on a 90-frame synthetic clip found that the choice of measurement region changed one matte's MSE from 1.41 to 9.03. The author argues that reports must state the region each metric covers, how it is normalized, and where the ground truth came from. Band-only reporting also hid over half of a chroma key's error, which sat outside the edge band. A control with pure flicker scored best on whole-frame dtSSD because its noise was diluted by the full-frame divisor. The author also notes unit scaling differences between raw JSON and the table, and that the dtSSD code was written from memory and not yet verified.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in