MECO: Mixture-of-Expert Codebooks for Multiple Dense Prediction Tasks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Autonomous systems require efficient scene understanding under computational constraints.
• Multi-task learning (MTL) offers a compact solution but faces challenges with entangled features and high computational overhead.

Purpose of the Study:

• To introduce Mixture-of-Expert Codebooks (MECO), a novel MTL framework designed to disentangle representations and reduce computational load.
• To enable efficient joint learning of dense prediction tasks for embedded systems.

Main Methods:

• Leveraging vector quantization to create lightweight Mixture-of-Experts (MoE) with disentangled task-generic and task-specific features.
• End-to-end training with a composite loss combining task-specific objectives and vector quantization losses.
• Evaluating MECO on semantic segmentation and monocular depth estimation tasks.

Main Results:

• MECO achieved a +0.4% mIoU improvement in semantic segmentation.
• Comparable accuracy in depth estimation compared to baseline models.
• Reduced model parameters by 18.33% and FLOPs by 28.83%.

Conclusions:

• MECO demonstrates the effectiveness of vector quantization-based MoE for efficient and scalable multi-task learning in embedded environments.
• The framework successfully addresses limitations of existing MTL approaches, offering a promising direction for resource-constrained autonomous systems.