Mobile-Seed is a lightweight, dual-task framework designed for simultaneous semantic segmentation and boundary detection. It’s essential for numerous robotic tasks such as robot grasping, real-time semantic mapping, and sensor calibration. The framework features a two-stream encoder and an active fusion decoder (AFD). One pathway captures category-aware semantic information while the other discerns boundaries from multi-scale features. The AFD module dynamically adapts the fusion of semantic and boundary information, allowing for precise weight assignment of each channel. This method has shown notable improvement over the state-of-the-art (SOTA) baseline in mIoU and mF-score, while maintaining a high online inference speed.
Publication date: 22 Nov 2023
Project Page: https://whu-usi3dv.github.io/Mobile-Seed/
Paper: https://arxiv.org/pdf/2311.12651
