This chapter introduces WSSIC-Net, a weakly-supervised approach for semantic instance completion. It reconstructs complete 3D objects from partial 2.5D scans without requiring full 3D supervision. This method addresses a significant limitation in current 3D scene understanding. Fully-supervised techniques often rely on expensive complete 3D annotations, which hinders practical deployment. The research makes three key contributions. First, it shows that semantic instance completion can be achieved using only bounding box annotations and unpaired synthetic data. Second, it presents a novel dual-branch architecture. This architecture combines supervised detection with weakly-supervised completion. Third, it demonstrates that competitive performance can be attained with significantly reduced annotation requirements. The framework consists of two synergistic components. The first is a fully-supervised 3D detection branch for proposal generation. The second is a weakly-supervised completion branch that uses adversarial learning with synthetic references. This design enables effective knowledge transfer from synthetic complete shapes to real-world partial scans while maintaining semantic consistency. Validation on ScanNet v2 shows that our method matches or exceeds fully-supervised approaches. This proves that weakly-supervised learning can be equally effective while being more practical. This chapter represents a significant step toward scalable 3D scene reconstruction by reducing dependency on costly annotations.