Simultaneous object detection an‎d semantic segmentation using Multi-task learning

10/22/2026 12:00:00 AM

Scene understan‎ding—particularly object detection an‎d semantic segmentation—is a fundamental task in domains such as autonomous driving an‎d robotics. Previous models that performed these two tasks simultaneously often struggled to balance accuracy an‎d computational complexity. This research introduces a novel an‎d efficient multi-task learning framework that addresses these challenges through three key innovations. The proposed method is built upon an integrated architecture featuring the following improvements: (1) replacing conventional convolution-based backbones with the advanced Swin Transformer architecture, which, due to its ability to capture global contextual information, leads to more precise feature representations; (2) employing the Low-Rank Adaptation (LoRA) technique to achieve an optimal trade-off between parameter count an‎d model accuracy, substantially reducing trainable parameters an‎d computational cost; an‎d (3) introducing a dynamic an‎d self-adaptive multi-task loss function that stabilizes the learning process through intelligent balancing among tasks. In addition, an edge detection task is incorporated as an auxiliary objective to enhance the precision of object boundaries. Experimental results on the PASCAL Context dataset demonstrate that the proposed architecture achieves a superior balance between model accuracy an‎d parameter efficiency while incorporating the additional edge detection task. Utilizing the LoRA technique, the model attains competitive performance with only 2.5 million trainable parameters (approximately 15× fewer than the reference paper). The model achieves 74.0% mAP for object detection an‎d 75.0% mIoU for semantic segmentation, comparable to the reference benchmark. These findings highlight a practical solution for deploying efficient scene understan‎ding systems on resource-constrained platforms such as embedded automotive processors.

يادگيري چند-وظيفه‌اي , تشخيص اشيا , بخش‌بندي معنايي , مبدل Swin , تطبيق با رتبه پايين

Multi-Task Learning , Object Detection , Semantic Segmentation , Swin Transformer , Low-Rank Adaptation

Ali Sheikhi

Dr. S. Ali Asghar Beheshti Shirazi