Prior-Guided Amodal Mask Completion: Revisiting Late-Stage Spatial Attention in Shape Reconstruction
Abstract
This paper investigates amodal mask completion under a prior-guided setting, isolating shape reconstruction from upstream detection errors using oracle visible masks and semantic labels. We propose a lightweight 5-channel Swin-UNet that fuses RGB, visible, and edge priors, optimized via a novel Occlusion-Aware Loss to prioritize invisible regions. Evaluated on KINS and COCOA, our framework achieves competitive accuracy while maintaining a compact architecture with only 34.42M parameters. Furthermore, systematic ablations yield key empirical insights: an intermediate occlusion weight (λ = 5) optimally balances overall reconstruction, whereas late-stage CBAM-style spatial attention counter-intuitively degrades occlusion hallucination by over-emphasizing local boundaries. These findings provide practical guidelines for designing efficient, prior-guided amodal segmentation systems.