Yanghao Li*, Yuntao Chen*, Naiyan Wang, Zhaoxiang Zhang

This repository implements TridentNet in the SimpleDet framework.

Trident Network (TridentNet) aims to generate scale-specific feature maps with a uniform representational power. We construct a parallel multi-branch architecture in which each branch shares the same transformation parameters but with different receptive fields. Then, we propose a scale-aware training scheme to specialize each branch by sampling object instances of proper scales for training. As a bonus, a fast approximation version of TridentNet could achieve significant improvements without any additional parameters and computational cost. On the COCO dataset, our TridentNet with ResNet-101 backbone achieves state-of-the-art single-model results by obtaining an mAP of 48.4.

• Dilated convolution for efficient scale enumeration
• Weight sharing between convs for uniform representation

The above figure shows how to convert bottleneck residual blocks to 3-branch Trident Blocks. The dilation rate of three branches are set as 1, 2 and 3, respectively.

Please setup SimpleDet following README and INSTALL and use the TridentNet configuration files in the config folder.

Note:

1. These models are not trained in SimpleDet. Re-training these models in SimpleDet gives a slightly better result.
2. TridentNet* - TridentNet = extended training + softNMS + multi-scale training/testing + syncBN + DCNv1.

All config files are available in config/resnet_v1b.

@article{li2019scale,
  title={Scale-Aware Trident Networks for Object Detection},
  author={Li, Yanghao and Chen, Yuntao and Wang, Naiyan and Zhang, Zhaoxiang},
  journal={ICCV 2019},
  year={2019}
}