Concretely, a U ^ 2-Net segmentation model imapanga per-pixel alpha matte, ndipo chiyambi chopanda kuwonongeka ndi chofunikira kwambiri: maziko oyera m'maziko, maziko oyera, ndi zotsatira zolemba zopanda kuwonongeka ndi kuwala kwabwino. There is no manual masking step — a neural segmentation model does the separation and you get a transparent result back.
Kodi ndi yotani yoyenera ndi cut-out pa mapiko ovuta?
+
Mafunso olimba m'mbuyomu amadziwika bwino. Maluwa abwino, maonekedwe, kuwala kwa motion, mawindo ndi masamba a chain-link ndi malo omwe machitidwe onse osinthika amalimbana, ndipo komwe mungafune kukhudza zotsatirazo ndi manja pambuyo pake.
Kodi mtundu wanga wa malemba amakhudza bwanji ubwino wa filimu yosaoneka bwino?
+
Yes. BMP ndi mwamsanga raw uncompressed RGB mnda ndi header ochepa, kotero ndi bwino lossless ndi zochuluka - nthawi zambiri 10-20x kukula kwa chithunzi chimodzimodzi monga PNG. Compression artefacts m'mawu a m'mawu ndi zomwe ziyenera kuchitidwa ndi segmentation model, kotero kulowa koyera kumapatsa chida choyera.
Pali china chilichonse chokhudzana ndi BMP chomwe ndingaganizire?
+
Yes — chifukwa palibe chomwe chimakonzedwa, BMP ndi kosavuta kuwerenga ndi kusintha koma kuwononga kutsitsa, ndipo sichibweretsa metadata kuti izikhala. Izi zimapanga zomwe model imawona komanso zomwe zitha kubweretsa zotsatira.
Yes — kutsitsa set ndi iwo kusonkhana mofanana pansi pa imodzi set of settings. Lossless owona ndi zikuluzikulu, kotero lalikulu gulu ndi kugwirizana ndi kutsitsa wanu galimoto kuposa ndi kusonkhana yekha.
Kodi alpha channel kupulumuka?
+
Ndiko. PNG ndi TIFF zimagwiritsa ntchito ma channels a 8-bit alpha ndipo zimasungana mpaka pamapeto. Zimakonzedwa ngati mwasankha mtundu womwe sumakhala ndi ma channels alpha, ndipo m'njira imeneyi ma pixels owoneka bwino amaphatikizidwa pa mthunzi wobiriwira.
