implementation of pp-matting v2

[AmberzzZZ]

问题确认 Search before asking

• [X] 我已经搜索过问题，但是没有找到解答。I have searched the question and found no related answer.

请提出你的问题 Please ask your question

hi, this is how you describe pp matting v2 in the front page: PP-MattingV2 is a lite matting SOTA model developed by PaddleSeg. It extracts high-level semantc informating by double-pyramid pool and spatial attention, and uses multi-level feature fusion mechanism for both semantic and detail prediciton. but in the code i see only a global alpha prediction head. How exactly does the semantic task and the detail task integrate into this network in training time.

[Asthestarsfalll]

I think the answer is in your question

It extracts high-level semantc informating by double-pyramid pool and spatial attention, and uses multi-level feature fusion mechanism for both semantic and detail prediciton.

Which corresponds to code

[AmberzzZZ]

Is the auxilliary semantic & detail prediction task no longer needed in this version of matting network?

[Asthestarsfalll]

Is the auxilliary semantic & detail prediction task no longer needed in this version of matting network?

I think double-pyramid pool, spatial attention and multi-level feature fusion grant the ppmatting-v2 the ability to extract high-level semantc information so that it doesn't need to use a auxilliary semantic map to supervise the training.

You may read the paper to obtain more details.