Journal of Applied Optics, Volume. 45, Issue 5, 946(2024)
Global-instance feature alignment domain adaptation detection method and system design
Figures & Tables(13)
Fig. 1. Block diagram of global-instance domain adaptation detection algorithm based on attention mechanism
Fig. 2. Schematic diagram of global feature alignment for background feature suppression
Fig. 3. Schematic diagram of image-level feature alignment and instance-level feature alignment
Fig. 4. Schematic diagram of multiple variable-length feature dictionaries
Fig. 5. Structure diagram of domain adaptation detection system design
Fig. 7. Visualization diagram of instance-level features in epoch 1, 10, 20, and 30 without/with instance-level feature alignment
Fig. 8. Detection results of model trained by Cityscapes on Foggy Cityscapes dataset
Table 1. Comparison of migration performance between proposed model and other domain adaptation methods from Cityscapes to Foggy Cityscapes dataset
View tableView in ArticleTable 1. Comparison of migration performance between proposed model and other domain adaptation methods from Cityscapes to Foggy Cityscapes dataset
方法 检测网络 Person Rider Car Truck Bus Train Motorcycle Bicycle mAP Source only Faster R-CNN 34.8 37.6 48.7 14.3 30.1 8.8 14.6 28.1 27.1 ConfMix[ 19 ]YOLOv5 45.0 43.4 62.6 27.3 45.8 40.0 28.6 33.5 40.8 SC-UDA[ 20 ]Faster R-CNN 38.5 43.7 56.0 27.1 43.8 29.7 31.2 39.5 38.7 MeGA-CDA[ 8 ]Faster R-CNN 37.7 49.0 52.4 25.4 49.2 46.9 34.5 39.0 41.8 MGA[ 21 ]Faster R-CNN 43.9 49.6 60.6 29.6 50.7 39.0 38.3 42.8 44.3 SSOD[ 22 ]Faster R-CNN 38.8 45.9 57.2 29.9 50.2 51.9 31.9 40.9 43.3 Ours Faster R-CNN 46.6 49.2 61.8 30.1 49.8 39.8 36.7 39.4 44.2 Oracle Faster R-CNN 52.2 55.0 72.1 31.2 52.7 45.0 33.9 50.0 49.0
Table 2. Ablation experiment of transferring learning from Cityscapes to Foggy Cityscapes dataset
View tableView in ArticleTable 2. Ablation experiment of transferring learning from Cityscapes to Foggy Cityscapes dataset
方法 注意力（损失） 注意力（特征） 字典学习 mAP50 性能提升 Baseline 27.1 Baseline √ 38.5 +11.4 Baseline √ 39.6 +12.5 Baseline √ 40.2 +13.1 Baseline √ √ 44.3 +17.2
Table 3. Comparison of detection performance before and after domain adaptation on self-owned datasets
View tableView in ArticleTable 3. Comparison of detection performance before and after domain adaptation on self-owned datasets
方法 骨干网络 mAP 不进行迁移 C2fDarknet 36.5 本文方法 C2fDarknet 44.2 监督学习 C2fDarknet 46.3
Table 4. Detection performance and number of parameters of network before and after using variable group convolution
View tableView in ArticleTable 4. Detection performance and number of parameters of network before and after using variable group convolution
方法 骨干网络 mAP 参数量/M 推理时间/ms 不进行迁移 C2fDarknet 31.5 3.2 74.2 本文方法 C2fDarknet 41.8 3.2 74.2 本文方法 C2fDarknet +VarGNet 40.5 2.5 23.4
Tools
Get Citation
Copy Citation Text
Yuan LIU, Yaxin LOU, Ping ZHANG, Yifan YANG, Yawei LI, Lingfan WU, Hong ZHANG. Global-instance feature alignment domain adaptation detection method and system design[J]. Journal of Applied Optics, 2024, 45(5): 946
Paper Information
Category:
Received: Sep. 28, 2023
Accepted: --
Published Online: Dec. 20, 2024
The Author Email: ZHANG Hong (张弘)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20