Fig 1.
Overview of the proposed ERPN.
Fig 2.
The architecture of deconvolutional feature pyramid network.
Fig 3.
Diagram of the interspersed scales for novel anchor boxes.
Fig 4.
Comparison between the anchor boxes of RPN and ERPN.
Fig 5.
The flowchart of optimizing the SVM parameters with PSO.
Fig 6.
The PSO iteration number.
Fig 7.
The classification loss function in RPN.
Fig 8.
The novel coefficients of improved classification loss function in ERPN.
Table 1.
Data set information.
Table 2.
Parameters for PSO.
Table 3.
Parameters of Fast R-CNN.
Table 4.
Parameters of MR-CNN.
Table 5.
Parameters of ION.
Table 6.
Parameters of ERPN.
Table 7.
Parameters of Faster R-CNN.
Table 8.
Parameters of HyperNet.
Table 9.
Detection results on PASCAL VOC 2007 test set, the best AP of each object category and mAP are bold-faced.
Table 10.
Detection results on PASCAL VOC 2012 test set, the best AP of each object category and mAP are bold-faced.
Fig 9.
Small objects detection results on PASCAL VOC 2007 and VCO 2012 test sets.
Fig 10.
Recall versus IoU threshold on the PASCAL VOC 2007 test set.
Left: 200 region proposals. Middle: 500 region proposals. Right: 1000 region proposals.
Table 11.
Detection results on MS COCO test-std.
The best result is bold-faced. A: improved anchor boxes, D: DFPN.
Table 12.
Comparison between SVM and softmax classifiers.
Table 13.
Experiment results for improved loss function.
Fig 11.
The influence for Eta on coefficients of improved classification loss function.
Table 14.
Detection speed of different methods on the PASCAL VOC 2007 test set.