HVM_BC

Video human matting is necessary for accurately predicting foreground human alpha mattes for each frame in a given input video sequence. Recent studies in this domain have mostly focused on trimap-free methods, as trimap-based methods are less adaptable for high-resolution real-time applications. Even though the latest trimap-free methods yield promising results, their performance often degrades when handling videos featuring overlapping foreground and background feature distributions. To address this shortcoming, we propose a novel method named High-resolution Real-time Video Human Matting Based on Bidirectional Consistency (HVM-BC), which leverages bidirectional temporal coherence and bidirectional decoding consistency in video sequences to establish the spatiotemporal relationships between the foreground and the background. The developed bidirectional recurrent decoder is incorporated into the HVM-BC, allowing not only the forward but also the backward temporal coherence in video sequences to be explored, providing comprehensive temporal context information for video human matting. In addition, we present a bidirectional decoding consistency training strategy that enhances the model’s attention on challenging regions where bidirectional decoding errors coincide, thereby yielding robust representations capable of identifying humans even in scenarios characterized by overlapping foreground and background feature distributions. Extensive experimental results demonstrate that HVM-BC outperforms state-of-the-art methods in terms of alpha matte quality and temporal coherence, particularly in cases where foreground and background feature distributions overlap.