Weakly Supervised Object Detection Based on Feature Self-Distillation Mechanism

Wenlong Gao; Ying Chen; Yong Peng

doi:10.3788/LOP212868

Laser & Optoelectronics Progress, Volume. 60, Issue 4, 0410009(2023)

Weakly Supervised Object Detection Based on Feature Self-Distillation Mechanism

Wenlong Gao, Ying Chen^*, and Yong Peng

Key Laboratory of Advanced Process Control for Light Industry of Ministry of Education, School of Internet of Things Engineering, Jiangnan University, Wuxi 214122, Jiangsu, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(14)

Fig. 1. Overall structure of network

Download full size

Fig. 2. Network architecture during test period

Download full size

Fig. 3. Multi-instance learning (MIL) and regression branch

Download full size

Fig. 4. Visualization of features extracted from Baseline and FSD-Net conv5 layer. (a) Original input image; (b) features extracted from Baseline; (c) features extracted from FSD-Net

Download full size

Fig. 5. Visualization of prediction results of Baseline (columns 1, 3, and 5) and FSD-Net (columns 2, 4, and 6)

Download full size

Table 1. Contribution of each improved module to detection accuracy
View table
Table 1. Contribution of each improved module to detection accuracy
Improved supervision generation algorithm Balancing optimization loss Regression branch Feature self-distillation mAP /%
－－－ 46.5
+ －－－ 49.2
+ + －－ 51.0
+ + + － 52.6
+ + + + 54.8

Table 2. Influence of feature self-distillation layers
View table
Table 2. Influence of feature self-distillation layers
Layer - conv 2+conv 3 conv 2+conv 4 conv 2+conv 3+conv 4
mAP /% 52.6 53.4 52.9 54.8

Table 3. Influence of threshold on detection accuracy in supervision information generation algorithm
View table
Table 3. Influence of threshold on detection accuracy in supervision information generation algorithm
Layer 0.00 0.05 0.10 0.15 0.20
mAP /% 52.5 55.3 54.8 52.7 48.2

Table 4. Influence of weight on detection accuracy in balancing optimization loss
View table
Table 4. Influence of weight on detection accuracy in balancing optimization loss
Weight $w$ - 2.0 2.5 3.0 3.5
mAP /% 52.6 52.8 54.5 55.3 53.0

Table 5. Detection performance of each category on VOC 2007 test set

View table

Table 5. Detection performance of each category on VOC 2007 test set

Method	OICR^［12］	WSCDN^［22］	MGR^［23］	CMIL^［13］	WSOD2^［24］	CMIDN^［25］	B-OICR^［15］	OIM+IR^［14］	FDC^［26］	FSD-Net
Aero	60.6	61.2	55.2	62.5	65.1	53.3	68.6	55.6	61.7	61.0
Bicycle	67.1	66.6	66.5	58.4	64.8	71.5	62.4	67.0	72.3	76.6
Bird	44.3	48.3	40.1	49.5	57.2	49.8	55.5	45.8	50.1	51.9
Boat	24.5	26.0	31.1	32.1	39.2	26.1	27.2	27.9	23.9	34.9
Bottle	19.2	15.8	16.9	19.8	24.3	20.3	21.4	21.1	9.1	28.0
Bus	68.9	66.5	69.8	70.5	69.8	70.3	71.1	69.0	70.9	74.0
Car	65.9	65.4	64.3	66.1	66.2	69.9	71.6	68.3	67.8	70.7
Cat	55.8	53.9	67.8	63.4	61.0	68.3	56.7	70.5	56.7	74.6
Chair	25.7	24.7	27.8	20.0	29.8	28.7	24.7	21.3	6.1	29.1
Cow	49.7	61.2	52.9	60.5	64.6	65.3	60.3	60.2	56.6	70.7
Table	43.7	46.2	47.0	52.9	42.5	45.1	47.4	40.3	40.8	49.5
Dog	47.4	53.5	33.0	53.5	60.1	64.6	56.1	54.5	61.8	65.6
Horse	33.8	48.5	60.8	57.4	71.2	58.0	46.4	56.5	55.8	60.5
Mbike	66.7	66.1	64.4	68.9	70.7	71.2	69.2	70.1	66.0	74.8
Person	10.6	12.1	13.8	8.4	21.9	20.0	2.7	12.5	3.3	16.0
Plant	24.8	22.0	26.0	24.6	28.1	27.5	22.9	25.0	20.6	25.3
Sheep	38.0	49.2	44.0	51.8	58.6	54.9	41.5	52.9	48.2	54.7
Sofa	52.5	53.2	55.7	58.7	59.7	54.9	47.7	55.2	60.5	63.9
Train	64.5	66.2	68.9	66.7	52.2	69.4	71.1	65.0	59.7	72.0
TV	66.2	59.4	65.5	63.5	64.8	63.5	69.8	63.7	58.0	52.2

Table 6. Localization performance of each category on VOC 2007 trainval set

View table

Table 6. Localization performance of each category on VOC 2007 trainval set

Method	OICR^［12］	WSCDN^［22］	PGE^［27］	MGR^［23］	WSOD2^［24］	B-OICR^［15］	FDC^［26］	FSD-Net
Aero	81.7	85.8	85.5	81.7	87.1	86.7	83.6	80.8
Bicycle	80.4	80.4	79.6	81.2	80.0	73.3	86.8	88.6
Bird	48.7	73.0	68.1	58.9	74.8	72.4	64.6	64.3
Boat	49.5	42.6	55.1	54.3	60.1	55.3	35.9	54.8
Bottle	32.8	36.6	33.6	37.8	36.6	46.9	25.4	48.9
Bus	81.7	79.7	83.5	83.2	79.2	83.2	82.3	84.8
Car	85.4	82.8	83.1	86.2	83.8	87.5	87.5	85.0
Cat	40.1	66.0	78.5	77.0	70.6	64.5	65.6	81.7
Chair	40.6	34.1	42.7	42.1	43.5	44.6	18.4	44.8
Cow	79.5	78.1	79.8	83.6	88.4	76.7	78.0	86.3
Table	35.7	36.9	37.8	51.3	46.0	46.4	46.0	46.8
Dog	33.7	68.6	61.5	44.9	74.7	70.9	76.0	79.5
Horse	60.5	72.4	74.4	78.2	87.4	67.0	80.8	83.3
Mbike	88.8	91.6	88.6	90.8	90.8	88.0	90.6	94.0
Person	21.8	22.2	32.6	20.5	44.2	9.6	7.6	27.4
Plant	57.9	51.3	55.7	56.8	52.4	56.4	53.9	59.7
Sheep	76.3	79.4	77.9	74.2	81.4	69.1	70.8	76.3
Sofa	59.9	63.7	63.7	66.1	61.8	52.4	75.1	68.5
Train	75.3	74.5	78.4	81.0	67.7	79.8	77.4	81.4
TV	81.4	74.6	74.1	86.0	79.9	82.8	82.4	71.7

Table 7. Comparison with mainstream methods in Pascal VOC 2007 dataset

View table

Table 7. Comparison with mainstream methods in Pascal VOC 2007 dataset

Method	mAP /%	CorLoc /%
WSCDN^［22］	48.3	64.7
PGE^［27］	47.6	66.7
MGR^［23］	48.6	66.8
C-MIL^［13］	50.5	65.0
WSOD2^［24］	53.6	69.5
C-MIDN^［25］	52.6	68.7
CSC^［28］	43.0	62.2
B-OICR^［15］	49.7	65.7
OIM+IR^［14］	50.1	67.2
FDC^［26］	47.5	64.4
ICM^［29］	54.8	68.8
PG-PS^［16］	51.1	69.2
MGML^［30］	53.0	67.1
DPS^［17］	50.9	66.5
GradingNet-MELM^［31］	52.5	63.2
OICR^［12］（Baseline）	46.5	60.6
FSD-Net	55.3	70.4

Table 8. Comparison with mainstream methods in Pascal VOC 2012 dataset

View table

Table 8. Comparison with mainstream methods in Pascal VOC 2012 dataset

Method	mAP /%	CorLoc /%
WSCDN^［22］	43.3	65.2
PGE^［27］	43.4	66.7
C-MIL^［13］	46.6	67.4
CSC^［28］	37.1	61.4
FDC^［26］	44.2	65.1
B-OICR^［15］	46.7	66.3
OIM+IR^［14］	45.3	67.1
PG-PS^［16］	48.3	68.7
MGML^［30］	48.5	67.4
DPS^［17］	43.8	-
GradingNet-MELM^［31］	48.6	62.8
OICR^［12］（Baseline）	37.9	62.1
FSD-Net	48.7	68.0

Table 9. Comparison with mainstream methods in MS-COCO dataset
View table
Table 9. Comparison with mainstream methods in MS-COCO dataset
Method AP₅₀ /%
WSCDN^［22］ 11.5
MELM^［32］ 18.8
PCL^［33］ 19.4
WS-JDS^［18］ 20.3
C-MIDN^［25］ 21.4
CSC^［28］ 20.3
PG-PS^［16］ 20.7
GradingNet-MELM^［31］ 22.6
OICR^［12］（Baseline） 14.3
FSD-Net 22.7

Tools

Get Citation

Copy Citation Text

Wenlong Gao, Ying Chen, Yong Peng. Weakly Supervised Object Detection Based on Feature Self-Distillation Mechanism[J]. Laser & Optoelectronics Progress, 2023, 60(4): 0410009

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites