Multi-object Detector with Double Task Network for Traffic Scenes

Enhancing the precision of multi-object detection within traffic scenes is of paramount importance. Despite significant advancements in object detection algorithms based on deep neural networks, the persistent challenges of inaccuracies and low efficiency in multi-object detection demand innovative solutions. We introduce an efficient object detection method, referred to as DTNet, aimed at elevating accuracy in multi-object scenarios and extending its utility to diverse object recognition tasks, including object detection and instance segmentation. To address these challenges, we propose DTNet, which incorporates three key components: the OPSM, the PACT, and the B-NMS. Our study demonstrates substantial performance enhancements compared to state-of-the-art YOLO detectors. When compared to YOLOv7-segment, DTNet achieves an 1.7% improvement in P on the Cityscapes-trf dataset. Furthermore, on the COCOtrf 17 dataset, DTNet exhibits an 50.42% enhancement in mAP@0.5, underscoring its efficacy in multi-object detection. The experimental outcomes demonstrate that our strategies not only enhance performance related to presicion, maintaining an equivalent of 142 GFLOPs, but also achieve this with a mere addition of 0.02 (M) in Params.