Parameter-Aware Mamba Model for Multitask Dense Prediction.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Multitask dense prediction requires understanding complex task inter-relations.
• Current methods often use convolutional layers and attention mechanisms.
• Transformers model holistic task relationships but can be computationally intensive.

Purpose of the Study:

• Introduce a novel decoder-based framework, the parameter-aware Mamba model (PAMM), for multitask dense prediction.
• Leverage state-space models (SSMs) for enhanced task interconnectivity and parameter efficiency.
• Improve the modeling of intrinsic task properties and global prior integration.

Main Methods:

• Developed a parameter-aware Mamba model (PAMM) utilizing dual state-space parameter experts (PEs).
• Integrated task-specific parameter priors (PPs) within the SSM framework.
• Employed multidirectional Hilbert scanning (MDHS) for multiangle feature sequence construction.

Main Results:

• PAMM demonstrated effectiveness in enhancing multitask dense prediction.
• The proposed method achieved strong performance on NYUD-v2 and PASCAL-Context benchmarks.
• The framework successfully integrated task-specific priors and improved feature representation.

Conclusions:

• PAMM offers a novel and effective approach for multitask dense prediction.
• SSMs provide a scalable and efficient alternative for modeling task interactions.
• The integration of PPs and MDHS enhances the model's ability to capture complex relationships.