This paper proposes a novel Trident Network (TridentNet) for object detection to address the challenge of scale variation. The key idea is to generate scale-specific feature maps with uniform representational power through a parallel multi-branch architecture. Each branch shares the same transformation parameters but has different receptive fields. A scale-aware training scheme is adopted to specialize each branch by sampling object instances of proper scales for training. This approach allows TridentNet to achieve state-of-the-art results on the COCO dataset, with a single model achieving 48.4 mAP using a ResNet-101 backbone. A fast approximation of TridentNet, TridentNet Fast, is also proposed, which achieves significant improvements without additional parameters or computational cost. The TridentNet structure is efficient and effective, with the ability to handle different scales of objects through parallel branches. The method is validated through extensive experiments, showing that TridentNet outperforms existing methods in terms of detection accuracy and efficiency. The proposed approach is also compatible with other methods like deformable convolution, which can adaptively adjust the receptive field. The results demonstrate that TridentNet is a promising solution for object detection in practical applications.This paper proposes a novel Trident Network (TridentNet) for object detection to address the challenge of scale variation. The key idea is to generate scale-specific feature maps with uniform representational power through a parallel multi-branch architecture. Each branch shares the same transformation parameters but has different receptive fields. A scale-aware training scheme is adopted to specialize each branch by sampling object instances of proper scales for training. This approach allows TridentNet to achieve state-of-the-art results on the COCO dataset, with a single model achieving 48.4 mAP using a ResNet-101 backbone. A fast approximation of TridentNet, TridentNet Fast, is also proposed, which achieves significant improvements without additional parameters or computational cost. The TridentNet structure is efficient and effective, with the ability to handle different scales of objects through parallel branches. The method is validated through extensive experiments, showing that TridentNet outperforms existing methods in terms of detection accuracy and efficiency. The proposed approach is also compatible with other methods like deformable convolution, which can adaptively adjust the receptive field. The results demonstrate that TridentNet is a promising solution for object detection in practical applications.