Boosting Video Object Segmentation via Robust and Efficient Memory Network

Recently, memory-based methods have exhibited remarkable performance in Video Object Segmentation (VOS) by employing non-local pixel-wise matching between the query and memory. Nevertheless, these methods suffer from two limitations: 1) Non-local pixel-wise matching can result in the incorrect segmentation of background distractor objects, and 2) memory features with substantial temporal redundancy consume significant computing resources and reduce the inference speed. To address the limitations, we first propose a local attention mechanism to suppress background features, and we introduce a novel training framework based on contrast learning to ensure the network learns reliable and robust pixel-wise correspondence between query and memory. We adaptively determine whether to update the memory based on the variation of foreground objects. Next, we propose a dynamic memory bank, which utilizes a lightweight and differentiable soft modulation gate to determine the number of memory features to remove along the temporal dimension. This allows efficient and flexible management of memory features. Our network achieves competitive results (e.g., 92.1% on DAVIS 2016 val, 87.6%/81.3% on DAVIS 2017 val/test, 87.0% on YouTube-VOS 2018 val) compared with the state-of-the-art methods while maintaining a faster inference speed of 25+FPS. Moreover, our network demonstrates a favorable balance between performance and speed when dealing with the long-time video dataset.