How to get same performance in image and video

[ToniButland1998]

When I test one video, I can get human and a green background.

But when I inference on images, I can only get a binary image, which is the mask.

How can I also get human with a green background, on image inference?

Thanks.