QIIS: Quaternion Inter- and Intra-Task Interaction Strategy for Multi-Task Dense Prediction

In the domain of intelligent consumer electronics, edge devices such as monocular cameras increasingly demand the ability to perform multiple visual perception tasks simultaneously, including semantic segmentation, depth estimation, and boundary detection. However, most existing multi-task dense prediction frameworks suffer from inefficient task interaction and limited feature integration, which significantly constrain their applicability in resource-constrained consumer environments. To address these limitations, we propose a Quaternion Inter- and Intra-task Interaction Strategy (QIIS), a unified framework designed to enhance task collaboration through quaternion-based feature fusion. Specifically, we introduce a Dual-stream Patch Embedding (DPE) module to extract rich local features, a Cross-prompt Quaternion Interaction (CQI) block to achieve efficient inter-task communication, and an Intra-Task Quaternion Fusion (ITQF) module that enables effective fusion of task-generic, task-specific, and cross-task representations. By leveraging the Hamilton product inherent in quaternion convolution, our approach enhances representational capacity while reducing parameter complexity. Experimental results on the PASCAL-Context and NYUD-v2 benchmarks demonstrate that QIIS achieves competitive performance across multiple tasks, with parameter and FLOPs counts that are generally lower than those of state-of-the-art methods, making it a viable solution for vision-based measurement and perception in embedded consumer systems. The code is available at https://github.com/hallynight/QIIS