RFD 1152: Background removal for photographic corpora
Problem
See-Through decomposes a character into semantic layers, and the union of its nineteen part masks is an obvious foreground matte. It is already in the manifest, already loaded, and on most frames it produces a clean cutout.
It is the wrong instrument for a photographic corpus, for a structural reason rather than a quality one. A union of semantic classes can only keep what it can name, and the model is trained on anime characters. Measured on a cosplay photograph, a Santa hat scored -8.38 on headwear – a confident negative, not a borderline miss – so the union placed a real garment in the background and deleted it. No threshold recovers it; at a threshold low enough to matter the union grows from 0.50 to 0.56 of the frame while headwear stays empty.
The same measurement showed the clothing classes fire on body region rather than on cloth: a nude torso is labelled topwear, and mask-derived coverage reads 0.81 to 0.93 on images that are largely bare skin.
Decision
Background removal uses a photo-trained dichotomous matting model, not a union of semantic part masks. ZhengPeng7/BiRefNet_HR-matting is the default. It has no vocabulary to fall outside of, so it cannot lose a garment for not knowing what the garment is.
See-Through is retained for what transfers – geometry – and not for what does not – semantics. Its part masks may be used as regions, never as labels for what covers what.
DETAILS.md carries the ground-truth ranking and the dtype it depends on.