Adversarial Active Learning Method for Enhanced Object Detection Guided by Uncertainty

Shaiya Wang; Hongzhong Tang; Haifan Luo; Shang Gao; Yuebing Xu

doi:10.3788/LOP240632

Laser & Optoelectronics Progress, Volume. 61, Issue 22, 2237002(2024)

Adversarial Active Learning Method for Enhanced Object Detection Guided by Uncertainty

Shaiya Wang¹, Hongzhong Tang^1,2、*, Haifan Luo¹, Shang Gao¹, and Yuebing Xu³

Author Affiliations

¹College of Automation and Electronic Information, Xiangtan University, Xiangtan 411105, Hunan , China

²The Key Laboratory of Intelligent Computing and Information Processing of Ministry of Education, Xiangtan University, Xiangtan 411105, Hunan , China

³College of Physics and Electronic Engineering, Hengyang Normal University, Hengyang 421002, Hunan , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(17)

Fig. 1. Sampling strategies for active learning. (a) Sampling based on uncertainty; (b) sampling based on diversity; (c) sampling based on mixed strategies; (d) proposed method

Download full size

Fig. 2. Architecture of AALGU model

Download full size

Fig. 3. Supervised Training

Download full size

Fig. 4. Kernel density plot of $l_{r e v}$

Download full size

Fig. 5. Graph of the uncertainty function

Download full size

Fig. 6. Unsupervised training

Download full size

Fig. 7. Experiments based on RetinaNet on MS COCO

Download full size

Fig. 8. Experiments based on Faster R-CNN on MS COCO

Download full size

Fig. 9. Experiments based on SSD on Pascal VOC 2007

Download full size

Fig. 10. Experiments based on SSD on Pascal VOC 07+12

Download full size

Fig. 11. Experiments with different r

Download full size

Fig. 12. Experiments with different T

Download full size

Fig. 13. Query sample distribution visualization

Download full size

Fig. 14. Categorical distribution map

Download full size

Table 1. Experiments based on Faster R-CNN on Pascal VOC 2007
View table
Table 1. Experiments based on Faster R-CNN on Pascal VOC 2007
Method 10% 15% 20% 25% 30% 35% 40% 45% 50%
Random 14.49 19.83 22.41 25.26 27.34 28.70 30.05 31.01 32.17
CALD 14.49 19.13 24.00 26.42 26.78 30.07 31.24 31.97 33.20
AALGU 14.49 23.72 26.49 28.01 29.46 30.74 31.27 32.58 33.51

Table 2. Experiments of various parts of AALGU

View table

Table 2. Experiments of various parts of AALGU

Method	UT	ST	RRS	10%	15%	20%	25%	30%	35%	40%	45%	50%
random				40.71	40.38	51.61	54.59	58.18	59.45	62.71	62.81	61.88
AALGU	√			40.71	45.89	47.20	57.29	60.76	61.09	61.22	62.25	62.51
	√		√	40.71	55.43	58.46	64.68	67.15	67.70	68.20	69.47	69.48
	√	√		40.71	60.29	62.23	65.35	66.83	67.96	68.37	68.74	69.24
	√	√	√	40.71	58.77	63.99	64.93	67.48	67.93	68.90	68.95	69.83

Table 3. Experiments of different training methods of adversarial networks
View table
Table 3. Experiments of different training methods of adversarial networks
Method 10% 15% 20% 25% 30% 35% 40% 45% 50%
1VAE+1D 40.71 60.22 60.90 64.10 66.31 68.85 68.66 68.73 69.19
1VAE+2D 40.71 57.89 62.90 65.61 66.96 67.29 68.49 68.78 69.51
2VAE+1D 40.71 58.77 63.99 64.93 67.48 67.93 68.90 68.95 69.83

Tools

Get Citation

Copy Citation Text

Shaiya Wang, Hongzhong Tang, Haifan Luo, Shang Gao, Yuebing Xu. Adversarial Active Learning Method for Enhanced Object Detection Guided by Uncertainty[J]. Laser & Optoelectronics Progress, 2024, 61(22): 2237002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites