<p>This paper addresses the task of semi-supervised video object segmentation and tracking, which aims to accurately segment target objects specified in the first frame throughout a video sequence. While current methods often rely on end-to-end networks trained extensively on large labeled datasets, their fixed weights can limit adaptability in complex scenarios. To enhance robustness, we propose three key innovations to the LWL baseline: (1) a multi-scale mask encoding strategy for richer target representation; (2) the integration of a local self-attention block into the segmentation decoder to improve feature discrimination; and (3) an adaptive weighting mechanism during offline training that prioritizes challenging samples to boost generalization. Extensive experiments on the DAVIS 2017 and VOT 2021 benchmarks demonstrate that our approach achieves state-of-the-art performance. In future work, we plan to incorporate target trajectory information to further improve segmentation and tracking accuracy.</p>