Add ViTMatte model

[amyeroberts]

Model description

ViTMatte is a recently released model for alpha matting on images i.e. background removal.

The model accepts an input image and trimap (manually labelled grayscale image outlining the rough border of the foreground object) and predicts the alpha mate for each pixel.

It introduces a series of small adaptations to the ViT architecture - selective global attention + window attention; adding convolutional blocks between transformers blocks - to reduce computational complexity and enhancing the high-frequency information passed through the network.

At the time of publishing, ViTMatte showed SOTA performance on Distinctions-646 and strong performance (> Mattformer) on Composition-1K.

Open source status

• The model implementation is available
• The model weights are available

Provide useful links for the implementation

Github: https://github.com/hustvl/ViTMatte
Paper: https://arxiv.org/pdf/2305.15272.pdf
Demo: https://colab.research.google.com/drive/1Dc2qoJueNZQyrTU19sIcrPyRDmvuMTF3?usp=sharing

[NielsRogge]

This can be closed now thanks to #25843