We use a two-stage Faster R-CNN style object detection framework [12] and leverage an SE-ResNeXt or SENet [4] model as the backbone feature extractor. To increase the global context information in the extracted features, we add an FPN and a pyramid spatial pooling (PSP) [15] module to the backbone. Additionally, we increase the context information in the head network by concatenating features from twice the area around each RoI to the head before the fullyconnected layers [17]. We increase the number of scales of features extracted by the feature extractor to five from four, which is used in the original work of FPN [7]. This modification allows the network to gather even greater global context information.

We use non-maximum weighted (NMW) [16] suppression during test time to reduce duplicate detections. It was found that this works better than standard non-maximum suppression (NMS). NMS was used in the RPN while training.

[15] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia. Pyramid scene parsing network. In CVPR, 2017.
[16] H. Zhou, Z. Li, C. Ning, and J. Tang. Cad: Scale invariant framework for real-time object detection. In ICCV Workshops, 2017.

[17] Y. Zhu, C. Zhao, J.Wang, X. Zhao, Y.Wu, H. Lu, et al. Couplenet: Coupling global structure with local parts for object detection. In ICCV, 2017.