Object detection in foggy image based on Double-Head

方法	骨干网络	汽车	公交车	自行车	摩托车	行人	mAP
Yolov3	DarkNet53	61.58	23.20	23.02	43.84	30.32	36.39
ATSS^［23］	ResNet101	60.95	35.69	10.37	46.33	30.32	37.68
Dynamic R-CNN^［24］	ResNet101	61.93	41.32	32.20	37.38	46.08	43.78
Double-Head	ResNet101	63.00	51.53	24.29	41.70	43.22	44.74
Cascade R-CNN	ResNet101	63.33	43.37	26.43	38.03	43.16	42.86
FPVDNet	ResNet101	65.47	44.10	26.50	44.31	46.61	45.40
KODNet	ResNet101	64.81	47.22	28.65	47.35	45.25	46.66
本文方法	ResNet101	66.41	50.78	33.62	50.12	46.92	49.37

Table 2. Performance comparison of different algorithms in the S-KITTI dataset

View table

View in Article

Table 2. Performance comparison of different algorithms in the S-KITTI dataset

方法	骨干网络	AP	AP₇₅	AP₅₀	AP_S	AP_M	AP_L
FSAF^［26］	ResNet101	62.8	69.6	89.9	63.7	63.0	66.8
ATSS	ResNet101	64.8	74.4	89.8	61.5	64.6	70.2
PANet	ResNet101	64.4	72.9	90.1	64.6	64.2	69.0
Libra R-CNN	ResNet101	64.5	74.3	89.7	63.8	63.7	69.5
Dynamic R-CNN	ResNet101	65.9	75.4	90.2	62.9	66.5	70.8
Double Head	ResNet101	65.4	74.5	90.7	65.4	64.6	69.8
本文方法	ResNet101	66.7	77.1	90.9	65.6	66.2	71.4

Table 3. Performance comparison of different algorithms in the S-COCOval dataset

View table

View in Article

Table 3. Performance comparison of different algorithms in the S-COCOval dataset

方法	骨干网络	AP	AP₇₅	AP₅₀	AP_S	AP_M	AP_L
FSAF	ResNet101	52.5	58.4	77.6	37.5	47.1	64.3
ATSS	ResNet101	54.4	61.2	78.0	38.7	48.8	66.4
PANet	ResNet101	53.6	60.7	78.2	35.6	47.8	64.8
Libra R-CNN	ResNet101	54.1	61.4	77.8	39.2	48.1	65.3
Dynamic R-CNN	ResNet101	56.8	63.6	78.7	39.4	50.4	69.0
Double Head	ResNet101	56.2	63.2	79.8	38.7	50.5	67.2
本文方法	ResNet101	57.7	65.1	79.6	41.1	51.1	69.3

Table 4. Comparison of the parameters and AP values after adding different number of modules b
View table
View in Article
Table 4. Comparison of the parameters and AP values after adding different number of modules b
模块a 模块b 参数/M AP
√ 0 1.06 37.1
√ 1 1.87 37.9（+0.8）
√ 2 2.68 38.6（+1.5）
√ 3 3.49 39.1（+2.0）
√ 4 4.30 39.6（+2.5）
√ 5 5.11 39.6（+2.5）
√ 6 5.92 39.7（+2.6）

Table 5. Ablation study of proposed component
View table
View in Article
Table 5. Ablation study of proposed component
方法注意力模块 mAP 参数量/M
/ 51.5 -
SE 51.6 +0.41
Double-Head CA 52.2 +0.52
CBAM 52.7 +0.51
本文 53.6 +0.45

Table 6. Ablation study of proposed component
View table
View in Article
Table 6. Ablation study of proposed component
主干网络注意力先验权重 EDH mAP Δ
ResNet50 52.9 -
ResNet50 √ 53.7 +0.8
ResNet50 √ 55.3 +2.4
ResNet50 √ 54.3 +1.4
ResNet50 √ √ √ 56.5 +3.6

Table 7. Experimental results of the proposed method under different foggy conditions
View table
View in Article
Table 7. Experimental results of the proposed method under different foggy conditions
雾浓度 β AP AP₇₅ AP₅₀
轻度 0.1~0.2 64.2 74.1 88.2
中度 0.2~0.3 62.7 72.5 87.6
重度 0.3~0.5 59.4 69.6 85.9

Tools

Get Citation

Copy Citation Text

Ren-si LI, Yun-yu SHI, Xiang LIU, Xian TANG, Jing-wen ZHAO. Object detection in foggy image based on Double-Head[J]. Chinese Journal of Liquid Crystals and Displays, 2023, 38(12): 1717

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Research Articles

Received: Mar. 7, 2023

Accepted: --

Published Online: Mar. 7, 2024

The Author Email: Yun-yu SHI (yunyushi@sues.edu.cn)

DOI:10.37188/CJLCD.2023-0089

Topics

Table 1. Performance comparison of different algorithms in the RTTS dataset

Table 1. Performance comparison of different algorithms in the RTTS dataset

Table 2. Performance comparison of different algorithms in the S-KITTI dataset

Table 2. Performance comparison of different algorithms in the S-KITTI dataset

Table 3. Performance comparison of different algorithms in the S-COCOval dataset

Table 3. Performance comparison of different algorithms in the S-COCOval dataset

Table 4. Comparison of the parameters and AP values after adding different number of modules b

Table 4. Comparison of the parameters and AP values after adding different number of modules b

Table 5. Ablation study of proposed component

Table 5. Ablation study of proposed component

Table 6. Ablation study of proposed component

Table 6. Ablation study of proposed component

Table 7. Experimental results of the proposed method under different foggy conditions

Table 7. Experimental results of the proposed method under different foggy conditions