Abstract
Autonomous systems operating in embedded environments require robust scene understanding under computational constraints. Multi-task learning offers a compact alternative to deploying multiple task-specific models by jointly solving dense prediction tasks. However, recent MTL models often suffer from entangled shared feature representations and significant computational overhead. To address these limitations, we propose Mixture-of-Expert Codebooks (MECO), a novel multi-task learning framework that leverages vector quantization to design Mixture-of-Experts with lightweight codebooks. MECO disentangles task-generic and task-specific representations and enables efficient learning across multiple dense prediction tasks such as semantic segmentation and monocular depth estimation. The proposed multi-task learning model is trained end-to-end using a composite loss that combines task-specific objectives and vector quantization losses. We evaluate MECO on a real-world driving dataset collected in challenging embedded scenarios. MECO achieves a +0.4% mIoU improvement in semantic segmentation and maintains comparable depth estimation accuracy to the baseline, while reducing model parameters and FLOPs by 18.33% and 28.83%, respectively. These results demonstrate the potential of vector quantization-based Mixture-of-Experts modeling for efficient and scalable multi-task learning in embedded environments.