Abstract
In object detection tasks, scale variation is one of the major challenges affecting detection accuracy. Existing methods typically enhance the scale generalization of models through strategies through image pyramids, feature pyramid networks, or multi-scale training. However, these approaches either increase inference computational costs or have not previously established explicit associations between different scale inputs during training. We propose Dual-Scale Parallel Training (DSPT) as a novel training strategy for object detection models. This method introduces high-and-low-resolution homologous image branches for associative learning during training. By fusing the losses from both branches, DSPT enables shared-parameter, same-iteration cross-scale joint optimization strategy, thereby enhancing the model’s multi-scale perception and generalization ability without altering the network structure or inference efficiency. Experimental results demonstrate that DSPT achieves stable performance improvements across several popular datasets (MS COCO, PASCAL VOC, VisDrone, etc.) and two model series (YOLO and RTDETR), with the mAP@50 metric improving by up to 4%. In scale-aware evaluation experiments, DSPT outperforms conventional and multi-scale training methods in detection accuracy and generalization, and provides positive gains across all object sizes. This strategy is simple, general, and scalable, providing a new approach to training object detection models with promising practical applications.
Funding Information
  • Guangxi University of Science and Technology (2026GXNSFHA00640243)
  • Guangxi University of Science and Technology (2026GXNSFHA00640260)
  • Doctoral Fund of Guilin University of Aerospace Technology (KX202201101)
  • Guangxi Experts of Science and Technology for Enterprises (KH202201103)
  • Science and Technology Program of Suzhou (20220107-3)