Optics and Precision Engineering, Volume. 32, Issue 5, 727(2024)
Position-sensitive Transformer aerial image object detection model
Figures & Tables(12)
Table 1. Ablation experiment results on VisDrone test set
View tableView in ArticleTable 1. Ablation experiment results on VisDrone test set
方法 MSFF PSSA Loss APS APM APL AP Param/M 基线 - - - 13.8 36.8 47.5 24.7 41.30 √ - - 16.4 38.9 49.4 26.4 42.36 - √ - 15.0 37.6 48.7 25.8 41.45 - - √ 15.6 39.1 48.9 26.0 41.30 √ √ - 17.1 39.7 49.8 27.2 42.51 - √ √ 16.5 40.0 49.1 26.9 41.45 √ - √ 18.5 39.6 50.1 28.1 42.36 Ours √ √ √ 19.4 40.1 50.9 28.8 42.51
Table 2. Experimental results for different attention mechanisms and using multi-scale features
View tableView in ArticleTable 2. Experimental results for different attention mechanisms and using multi-scale features
组别 方 法 APS APM APL AP A Baseline 13.8 36.8 47.5 24.7 B Baseline-SE 13.9 37.0 47.5 24.9 C Baseline-SA 14.5 38.1 47.7 25.2 D Baseline-CA 14.3 37.7 48.3 25.4 E Baseline-CBAM 14.6 37.5 48.1 25.2 F Baseline-PCE3DA 15.2 38.4 48.7 25.7 G F+MSFF 16.4 38.9 49.4 26.4
Table 3. Experimental results of different relative position calculation methods
View tableView in ArticleTable 3. Experimental results of different relative position calculation methods
方 法 APS APM APL AP 基线模型 13.8 36.8 47.5 24.7 文献［ 27 ］14.3 37.0 48.3 25.0 文献［ 28 ］14.6 37.4 48.1 25.1 PSSA 15.0 37.6 48.7 25.8
Table 4. Performance comparison of different algorithms on VisDrone test set
View tableView in ArticleTable 4. Performance comparison of different algorithms on VisDrone test set
方 法 AP50 AP75 AP FPS Faster R-CNN［ 3 ］21.7 / / 15.9 Cascade R-CNN［ 4 ］38.6 25.0 23.5 9.0 YOLOv4［ 6 ］31.2 16.7 16.8 28.8 QueryDet［ 7 ］48.1 28.8 28.3 2.8 CornerNet［ 10 ］34.1 15.8 17.4 15.5 RetinaNet［ 20 ］28.4 12.3 11.3 16 Double-Head RCNN［ 29 ］38.3 24.8 23.8 6.5 IterDet［ 30 ］36.8 20.3 20.4 11.4 RSOD［ 31 ］43.3 27.1 25.4 28 YOLOv8［ 32 ］46.4 27.5 26.5 30.1 PVTv2［ 33 ］34.1 21.4 20.6 10.9 PS-TOD（Ours） 51.8 28.3 28.8 22.7
Table 5. Experimental results of different categories on VisDrone test set
View tableView in ArticleTable 5. Experimental results of different categories on VisDrone test set
目标类别 行人 人 汽车 公交车 自行车 卡车 三轮车 雨棚三轮车 面包车 摩托车 基线模型 24.8 18.7 61.6 35.2 12.1 23.3 15.2 4.6 28.6 24.9 PS-TOD 29.0 22.4 64.3 45.9 14.7 27.1 21.4 9.0 31.7 28.4
Tools
Get Citation
Copy Citation Text
Daxiang LI, Jiani XIN, Ying LIU. Position-sensitive Transformer aerial image object detection model[J]. Optics and Precision Engineering, 2024, 32(5): 727
Paper Information
Category:
Received: May. 30, 2023
Accepted: --
Published Online: Apr. 2, 2024
The Author Email: Jiani XIN (xjn_2000@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20