EfficientMatting: Bilateral Matting Network for Real-Time Human Matting

Recent human matting methods typically suffer from two drawbacks: 1) high computation overhead caused by multiple stages, and 2) limited practical application due to the need for auxiliary guidance (e.g., trimap, mask, or background). To address these issues, we propose EfficientMatting, a real-time human matting method using only a single image as input. Specifically, EfficientMatting incorporates a bilateral network composed of two complementary branches: a transformer-based context information branch and a CNN-based spatial information branch. Furthermore, we introduce three novel techniques to enhance model performance while maintaining high inference efficiency. Firstly, we design a Semantic Guided Fusion Module (SGFM), which empowers the model to dynamically acquire valuable features with the assistance of context information. Secondly, we design a lightweight Detail Preservation Module (DPM) to achieve detail preservation and mitigate image artifacts during the upsampling process. Thirdly, we introduce the Supervised-Enhanced Training Strategy (SETS) to explicitly provide supervision on hidden features. Extensive experiments on P3M-10k, Human-2K, and PPM-100 datasets show that EfficientMatting outperforms state-of-the-art real-time human matting methods in terms of both model performance and inference speed.