Shape Distribution Matters: Shape-Specific Mixture-of-Experts for Amodal Segmentation Under Diverse Occlusions
ShapeMoE is a shape-specific sparse Mixture-of-Experts framework for amodal segmentation, which predicts complete object masks including occluded areas. It learns a latent shape distribution to dynamically route objects to specialized, lightweight experts tailored to different shapes. This approach improves performance over existing methods by ensuring the right expert handles complex occlusions efficiently.