Portrait matting using an attention-based memory network

We propose a novel network to perform auxiliary-free video matting task. Unlike most existing approaches that require trimaps or pre-captured backgrounds as auxiliary inputs, our method uses binary segmentation masks as priors and realizes the auxiliary-free matting. Furthermore, we design the attention-based memory block by combining the idea of the memory network and self-attention to compute pixel-level temporal coherence among video frames to enhance the overall performance. Moreover, we also provide direct supervision for the temporal-guided memory module to boost the network’s robustness. The validation results on various testing datasets show that our method outperforms several state-of-the-art auxiliary-free matting methods in terms of the alpha and foreground prediction quality and temporal consistency.