Abstract | ||
---|---|---|
In this paper, we present rotational region-based fully convolutional networks (RR-FCN) for object detection. In contrast to previous detectors that do not consider rotation, our region-based detector incorporates rotational invariance into networks efficiently and generate more appropriate features according to the rotation angle. Specifically, we propose component-sensitive feature maps, rotational RoI pooling and interceptive back propagation which make RR-FCN learn rotation situations without extra supervision information. Using the 101-layer ResNet model, our method achieves state-of-the-art detection accuracy on PASCAL VOC 2007 and 2012. Moreover, since the feature maps in our network are component-sensitive, RR-FCN can find out objects with various postures, even those appear rarely in the training set. So our RR-FCN has better performance in the real world. |
Year | Venue | Field |
---|---|---|
2018 | EANN | Training set,Rotational invariance,Object detection,Computer science,Pooling,Algorithm,Residual neural network,Backpropagation,Detector |
DocType | Citations | PageRank |
Conference | 0 | 0.34 |
References | Authors | |
10 | 4 |
Name | Order | Citations | PageRank |
---|---|---|---|
Dingqian Zhang | 1 | 0 | 0.34 |
Hui Zhang | 2 | 104 | 7.12 |
Haichang Li | 3 | 0 | 0.34 |
Xiaohui Hu | 4 | 17 | 8.10 |