YOLOMP: Fast and Balanced Multi-Task Perception for Autonomous Driving

Engineering notes

Experiments on the BDD100K dataset demonstrate that YOLOMP achieves real-time inference at 6.9ms with up to 62.5% fewer parameters than YOLOP, while delivering competitive or superior accuracy compared to YOLOP and other existing multitask baselines.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Autonomous driving demands real-time perception that balances accuracy across multiple visual tasks under tight computational budgets. This study presents YOLOMP, a lightweight multi-task framework that integrates object detection, drivable-area segmentation, and lane-line segmentation within a unified model. Built on a YOLOv11-derived backbone with task-specific necks and decoders, YOLOMP maximizes feature sharing to reduce computational cost while preserving task-specific adaptability. To ensure balanced performance across multi-tasks, we adopt PCGrad, which alleviates inter-task gradient conflicts and prevents performance degradation in any single objective. Experiments on the BDD100K dataset demonstrate that YOLOMP achieves real-time inference at 6.9ms with up to 62.5% fewer parameters than YOLOP, while delivering competitive or superior accuracy compared to YOLOP and other existing multitask baselines. These results highlight YOLOMP's efficacy in delivering efficient, balanced multi-task performance for real-world resource-constrained autonomous driving deployments.

Links and sources

• Official / arXiv page

Comments