I can confirm this with a couple of Codepen demos. MediaPipe's selfie-segmentation model[1] seems to give much better results than TensorFlow's body-pix model[2] (which is the model I most often see talked about for this sort of background removal work).
As much as I admire ML models (and I remain in awe of people who know how to build them), I still think there's room for more conventional chroma key filtering - for instance identifying part of a video's moving display so you can add some UI to it[3].
As much as I admire ML models (and I remain in awe of people who know how to build them), I still think there's room for more conventional chroma key filtering - for instance identifying part of a video's moving display so you can add some UI to it[3].
[1] - MediaPipe demo - https://codepen.io/kaliedarik/pen/PopBxBM
[2] - TensorFlow demo - https://codepen.io/kaliedarik/pen/ZEeoZaP
[3] - https://scrawl-v8.rikweb.org.uk/demo/canvas-027.html