Improved Algorithm Based on Feature Pyramid Networks

Jingming Chen; Jie Jin; Weifeng Wang

doi:10.3788/LOP56.211505

Journals >Laser & Optoelectronics Progress >Volume 56 >Issue 21 >Page 211505 > Article

Laser & Optoelectronics Progress
Vol. 56, Issue 21, 211505 (2019)

Improved Algorithm Based on Feature Pyramid Networks

Jingming Chen, Jie Jin^*, and Weifeng Wang

Author Affiliations

School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China

show less

DOI: 10.3788/LOP56.211505 Cite this Article Set citation alerts

Jingming Chen, Jie Jin, Weifeng Wang. Improved Algorithm Based on Feature Pyramid Networks[J]. Laser & Optoelectronics Progress, 2019, 56(21): 211505 Copy Citation Text

show less

Fig. 1. Overall network structure of Refine-FPN algorithm

Download full size

Fig. 2. Basic structure of prediction optimization module

Download full size

Fig. 3. Test results

Download full size

Method	Training set	Backbone	Accuracy /%
Method	Training set	Backbone	IOU of 0.5	IOU of 0.6	IOU of 0.75
SSD	VOC07+12	VGG-16	77.3	72.3	61.3
RFCN	VOC07+12	ResNet-101	80.5	73.2	61.8
Faster Rcnn	VOC07+12	ResNet-101	76.4	69.5	57.3
YOLOv2	VOC07+12	Darknet-19	78.6	69.1	56.5
FPN	VOC07+12	ResNet-101	80.5	-	-
Couplenet	VOC07+12	ResNet-101	81.7	-	-
Res101-RFCN-Cascade	VOC07+12	ResNet-101	79.6	-	59.2
BPN512	VOC07+12	VGG-16	81.9	77.6	68.3
Refine-FPN	VOC07+12	Detnet-59	80.9	74.4	65.3
FPN^*	VOC07+12	Detnet-59	79.8	-	-

Table 1. Comparison between Refine-FPN and other classical algorithms

Method	mAP	Bike	Boat	Bottle	TV	Chair	Table	Sheep	Train	Cat
SSD	77.3	83.9	69.6	50.5	76.8	60.3	77.0	77.9	87.6	88.1
RFCN	80.5	89.6	69.0	69.2	79.5	65.4	72.1	79.6	87.1	88.4
Faster Rcnn	76.4	80.7	68.3	55.9	72.0	56.7	69.4	78.6	85.3	85.3
FPN	80.5	80.1	72.9	67.4	72.3	61.5	68.7	78.3	87.4	87.3
Couplenet	81.7	86.0	74.5	72.3	80.1	68.8	75.6	81.9	86.7	88.5
Refine-FPN	80.9	80.4	73.8	70.5	73.1	63.3	69.8	79.1	87.6	87.8

Table 2. Specific test results on VOC2007%

Stage	Accuracy (IOU of 0.5) /%
First stage (IOU of 0.5)	79.8
Second stage(IOU of 0.6)	80.3
Third stage(IOU of 0.75)	80.5
Fourth stage(IOU of 0.8)	80.4

Table 3. Comparison results when different prediction networks are cascaded in FPN^* algorithm

Context information	Accuracy (IOU of 0.5) /%
ROI(×1)	79.8
ROI(×0.8, ×1)	80.0
ROI(×1, ×1.2)	80.1
ROI(×0.8, ×1, ×1.2)	80.3

Table 4. Comparison results when different contextual information are combined in FPN^* algorithm

Jingming Chen, Jie Jin, Weifeng Wang. Improved Algorithm Based on Feature Pyramid Networks[J]. Laser & Optoelectronics Progress, 2019, 56(21): 211505

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information