Object Detection Based on Semantic Sampling and Localization Refinement

Yu Li; Shaoyan Gai; Feipeng Da; Ru Hong

doi:10.3788/LOP202259.1815015

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 18 >Page 1815015 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 18, 1815015 (2022)

Object Detection Based on Semantic Sampling and Localization Refinement

Yu Li^1、2, Shaoyan Gai^1、2、3, Feipeng Da^{1、2、3、*}, and Ru Hong^1、2

Author Affiliations

¹School of Automation, Southeast University, Nanjing 210096, Jiangsu, China

²Key Laboratory of Measurement and Control of Complex Systems of Engineering, Ministry of Education,Southeast University, Nanjing 210096, Jiangsu, China

³Shenzhen Research Institute, Southeast University, Shenzhen 518063, Guangdong, China

show less

DOI: 10.3788/LOP202259.1815015 Cite this Article Set citation alerts

Yu Li, Shaoyan Gai, Feipeng Da, Ru Hong. Object Detection Based on Semantic Sampling and Localization Refinement[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1815015 Copy Citation Text

show less

Fig. 1. Comparison of object detection architectures. (a) Architecture of RepPoints^[14]; (b) architecture of proposed detector

Download full size

Fig. 2. Overall architecture of proposed object detector. Dashed box: semantic based positioning module; dashed box: feature enhancement module

Download full size

Fig. 3. Architecture of localization subnet and classification subnet

Download full size

Fig. 4. Training samplers for object detection. (a) Center-based training sampler; (b) semantic-based training sampler

Download full size

Fig. 5. Qualitative results of proposed method on different datasets. (a), (b) VOC 2007; (c), (d) MS COCO

Download full size

Method	Backbone	AP	$A P_{50}$	$A P_{75}$	$A P_{s}$	$A P_{m}$	$A P_{l}$	Time /ms
RetinaNet^［22］	ResNet-101	39.1	59.1	42.3	21.8	42.7	50.2	90
FCOS^［6］	ResNet-101	41.5	60.7	45.0	24.4	44.8	51.6	91
FSAF^［5］	ResNet-101	40.9	61.5	44.0	24.0	44.2	51.3	138
ATSS^［10］	ResNet-101	43.6	62.1	47.4	26.1	47.0	53.6	93
RepPoints^［14］	ResNet-101	41.0	62.9	44.3	23.6	44.1	51.7	87
CornerNet^［4］	Hourglass-104	40.5	56.5	43.1	19.4	42.7	53.9	227
ExtremeNet^［7］	Hourglass-104	40.2	55.5	43.2	20.4	43.2	53.1	348
CenterNet^［8］	Hourglass-104	42.1	61.1	45.9	24.1	45.5	52.8	298
LSNet^［21］	ResNeXt-101	43.9	63.1	47.8	26.6	47.1	55.4	138
Proposed	ResNet-101	42.8	65.1	46.3	26.1	47.3	55.8	78

Table 1. Experimental result comparison of proposed method to other methods on MS COCO test-dev

Method	AP	$A P_{50}$	$A P_{75}$	$A P_{s}$	$A P_{m}$	$A P_{l}$	Time /ms
Baseline RepPoints	39.1	60.5	42.1	21.2	41.2	50.0	85
RepPoints + LB	40.4	61.9	43.6	23.1	43.8	52.4	85
RepPoints + LB + FE	41.0	62.6	44.3	24.3	44.6	53.3	86
RepPoints + EO	39.2	61.7	42.2	21.4	42.8	50.6	76
Proposed（LB + FE+EO）	41.2	63.0	44.2	24.6	45.0	53.9	77

Table 2. Experimental result comparison of integrating three modules into RepPoints

Conversion function $N_{L}$	AP	$A P_{50}$	$A P_{75}$	Time /ms
1×1 conv	40.4	60.2	43.6	85
Max-pooling	40.3	60.1	43.4	85
Max-pooling + avg-pooling	40.4	60.3	43.8	86

Table 3. Comparison of different transformation conversions

$ϵ_{L}$	AP	$A P_{50}$	$A P_{75}$
0.000	39.1	59.9	42.1
0.005	39.8	60.1	42.1
0.010	40.4	60.2	43.6
0.050	40.2	60.1	43.3
0.100	39.8	59.9	42.9

Table 4. Detection results with different location thresholds

Method	EO	FLOPs /G	Param /M	FPS /（frame·s^-1）	Time /ms	AP
RepPoints		190.16	36.62	11.8	85	39.1
RepPoints	✓	190.06	35.97	13.1	76	39.2
Proposed		190.18	36.62	11.6	86	41.0
Proposed	✓	190.07	35.97	12.9	77	41.2

Table 5. Impact of applying efficient refinement optimization to different networks

Yu Li, Shaoyan Gai, Feipeng Da, Ru Hong. Object Detection Based on Semantic Sampling and Localization Refinement[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1815015

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information