NAS-FPN为了提升搜索速度以及减少Two Stage目标检测的proposal采样以及RCNN部分的影响,选择了RetinaNet[4]作为搜索baseline,同样的,和RetinaNet保持一致,也采用了C3-C7进行特征融合。和分类网络架构搜索方法NASNet[5]类似,NAS-FPN也采用了以RNN作为控制器的强化学习搜索方法,定义了节点集合(可用的特征图节点集合),操作池(sum 和 global pooling)以及搜索终止条件(填满输出金字塔的每一层)。如图:
[1] Ghiasi, Golnaz, Tsung-Yi Lin, and Quoc V. Le. "Nas-fpn: Learning scalable feature pyramid architecture for object detection."Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019.
[2] Lin, Tsung-Yi, et al. "Feature pyramid networks for object detection."Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.
[3] Liu, Shu, et al. "Path aggregation network for instance segmentation." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.
[4]Lin, Tsung-Yi, et al. "Focal loss for dense object detection."Proceedings of the IEEE international conference on computer vision. 2017.
[5] B. Zoph, V. Vasudevan, J. Shlens, and Q. V. Le. Learning transferable architectures for scalable image recognition. InCVPR, 2018.
[6] H. Li, P. Xiong, J. An, and L. Wang. Pyramid attention net- work for semantic segmentation. BMVC, 2018.