Few-Shot Object Detection by Second-Order Pooling
Few-Shot Object Detection by Second-Order Pooling
复制标题
DOI：
10.1007/978-3-030-69538-5_23
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
Shan Zhang;Dawei Luo;Lei Wang-;Piotr Koniusz
中科院分区：
文献类型：
--
作者：
Shan Zhang;Dawei Luo;Lei Wang-;Piotr Koniusz
在本文中，我们解决了一个具有挑战性的问题，少镜头目标检测，而不是识别。提出了一种功率归一化二阶检测器，它由编码网络（EN）、多尺度特征融合（MFF）、功率归一化二阶池（SOP）、超注意区域建议网络（HARPN）和相似性网络（SN）组成。EN采取支持图像作物和查询图像每集产生跨几个层的卷积特征图，而MFF将它们组合成多尺度特征图。SOP根据支持图像聚合它们，而PN检测视觉特征的存在，而不是计算其发生频率。HARPN将PN汇集的支持特征与查询特征图进行交叉相关，以匹配区域并产生查询区域建议，然后将其与SOP/PN聚合。最后，支持和查询二阶描述符被传递给SN。我们的方法表现良好，因为：（i）HARPN利用SOP/PN进行检测到的而不是计数的支持特征与查询特征的互相关，这改善了区域建议，（ii）SOP/PN捕获每个区域建议的二阶统计量并分解出空间位置，以及（iii）PN限制了HARPN和SN学习的函数空间的复杂性。这些属性导致PASCAL VOC 2007/12，MS COCO和FSOD数据集的最新技术水平。
In this paper, we tackle a challenging problem of Few-shot Object Detection rather than recognition. We propose Power Normalizing Second-order Detector consisting of the Encoding Network (EN), the Multi-scale Feature Fusion (MFF), Second-order Pooling (SOP) with Power Normalization (PN), the Hyper Attention Region Proposal Network (HARPN) and Similarity Network (SN). EN takes support image crops and a query image per episode to produce covolutional feature maps across several layers while MFF combines them into multi-scale feature maps. SOP aggregates them per support image while PN detects the presence of visual feature instead of counting its frequency of occurrence. HARPN cross-correlates the PN pooled support features against the query feature map to match regions and produce query region proposals that are then aggregated with SOP/PN. Finally, support and query second-order descriptors are passed to SN. Our approach performs well because:(i) HARPN leverages SOP/PN for cross-correlation of detected rather than counted support features with query features which improves region proposals,(ii) SOP/PN capture second-order statistics per region proposal and factor out spatial locations, and (iii) PN limits the complexity of the space of functions over which HARPN and SN learn. These properties lead to the state of the art on the PASCAL VOC 2007/12, MS COCO and the FSOD datasets.