Researchers Explore Frequency-Aware Task Weighting for YOLOv8 Autonomous Driving

In a research preprint published on arXiv, researchers investigated Frequency-aware Task Weighting (FTW) for unified perception systems in autonomous driving. The approach integrates object detection, drivable-area segmentation, and lane segmentation within a single YOLOv8-based architecture.

FTW dynamically balances tasks by estimating a loss-trajectory stability proxy derived from the low-frequency energy ratio of recent loss histories. The method assigns greater weights to tasks with higher low-frequency power in their mean-centered loss trajectories. Testing was conducted on the Mapillary Vistas dataset under static full-network training and progressive freezing.

Single-run experiments indicated that FTW achieved top results in specific metrics, such as lane mIoU and detection mAP. However, the authors noted that without multi-seed evaluations or extensive ablations, the findings remain descriptive and validate feasibility rather than definitive superiority over baselines.