DPNeXt Decoder for Efficient Multi-Task Dense Prediction in Robotics

DPNeXt replaces the Dense Prediction Transformer (DPT) with a multi-scale feature fusion decoder using dual depthwise separable inverted bottlenecks. It incorporates Multi-Task Boundary Guidance (MTBG) to reduce negative inductive transfer between tasks like semantic segmentation and depth estimation without requiring additional annotations.