Abstract

Human memory, while remarkable, struggles to retain and recall intricate details from various experiences. This abstract proposes a novel “Experience Memory Network” to address this limitation in the context of video semantic segmentation. The network functions as a fully connected architecture, comprising information storage units that retain data on individual video frames and relations that capture the connections between them. To facilitate efficient information retrieval and updates, the network incorporates specialized read and write modules. This architecture empowers the system to leverage past experiences (stored in the memory network) to perform semantic segmentation on new video frames. Essentially, the network learns from each frame and utilizes this knowledge to understand and segment subsequent frames, leading to improved overall performance. The segmentation evaluation of Experience Memory Network yields two crucial metrics: mean region similarity (80.2%), indicating the overall overlap between predicted and true regions, and mean contour accuracy (84.2%), which specifically assesses how well predicted contours align with actual object boundaries. Analysing both metrics together provides a holistic understanding of segmentation effectiveness, encompassing both the extent of overlap and the precision of boundary delineation.