Acta Photonica Sinica, Volume. 50, Issue 10, 1010003(2021)
A Person Detection Algorithm in Fisheye Images Based on Rotated Boxes
Figures & Tables(13)
Fig. 2. Different strategies used as bounding box regression samples
Fig. 4. Comparison between fixed Gaussian kernel and adaptive Gaussian kernel
Fig. 5. Transformation from network regression outputs to predicted bounding boxes
Fig. 6. Minimum axis-aligned rectangle of two rotated bounding boxes
Fig. 7. Relationship between IoU and angle difference under different aspect ratios
Table 1. Comparative experiment of Gaussian kernel function
View tableView in ArticleTable 1. Comparative experiment of Gaussian kernel function
Gaussian kernel MW⁃R/% HABBOF/% CEPDOF/% mAP AP75 mAP AP75 mAP AP75 Axis-aligned Gaussian kernel 31.9 23.3 40.5 35.2 17.1 6.4 Rotated Gaussian kernel 31.7 24.4 42.7 37.2 18.3 8.9
Table 2. Comparative experiment of IoU loss
View tableView in ArticleTable 2. Comparative experiment of IoU loss
Loss type MW⁃R/% HABBOF/% CEPDOF/% mAP AP75 mAP AP75 mAP AP75 PIoU 43.6 33.9 49.2 47.9 27.8 14.8 GIoU 27.3 13.2 15.2 6.0 9.3 1.7 CIoU 42.3 31.7 49.5 46.8 26.2 14.8 EIoU 45.4 37.2 49.3 48.1 29.2 16.1 AIoU 46.4 40.3 50.0 49.4 29.6 17.0
Table 3. Ablation experiment of hyperparameter
View tableView in ArticleTable 3. Ablation experiment of hyperparameter
MW⁃R/% mAP AP50 AP75 1.0 45.4 90.1 37.9 1.3 45.2 90.3 36.3 1.6 45.5 90.2 38.5 1.9 46.4 90.9 40.3 2.2 46.4 90.4 39.6 2.5 45.8 90.6 37.3
Table 4. Comparative experiment of Gaussian kernel function
View tableView in ArticleTable 4. Comparative experiment of Gaussian kernel function
Loss type MW⁃R/% HABBOF/% CEPDOF/% mAP AP75 mAPs mAP AP75 mAPs mAP AP75 mAPs PIoU 43.6 33.9 34.3 49.2 47.9 54.7 27.8 14.8 18.3 PIoU_A 44.3 35.3 34.9 49.7 48.3 57.2 28.1 15.5 19.6 CIoU 42.3 31.7 33.0 49.5 46.8 53.4 26.2 14.8 15.9 CIoU_A 42.9 32.2 33.4 49.6 47.8 53.8 29.0 15.7 19.8 EIoU 45.4 37.2 33.4 49.3 48.1 53.7 29.2 16.1 19.0 EIoU_A 46.2 38.4 34.7 49.8 50.3 55.3 30.1 18.4 19.1 AIoU 46.4 40.3 36.7 50.0 49.4 55.8 29.6 17.0 19.7 AIoU_A 47.2 40.6 35.9 49.6 50.0 54.1 30.1 18.0 19.6
Table 5. Comparison of the state-of-the-art algorithm and proposed algorithm
View tableView in ArticleTable 5. Comparison of the state-of-the-art algorithm and proposed algorithm
Method Size FPS MW⁃R/% HABBOF/% CEPDOF/% mAP AP50 AP75 F mAP AP50 AP75 F mAP AP50 AP75 F RAPiD 608 20.5 53.4 96.8 50.9 94.1 56.7 96.7 62.6 95.8 38.2 83.1 27.6 79.3 RAPiD 1 024 13.6 53.3 96.7 48.8 93.5 57.3 97.8 59.1 96.9 39.4 86.4 26.7 83.6 RAPiD_2x 608 17.4 42.5 93.2 26.3 89.3 43.7 90.7 31.2 88.5 29.3 78.3 11.6 78.0 AFRPD-34 608 78.5 53.3 95.9 53.0 93.8 54.1 95.9 56.7 93.2 36.6 84.2 23.7 81.8 AFRPD-34 1 024 39.0 55.5 97.9 55.9 96.2 59.4 97.2 63.5 94.4 40.1 86.1 29.2 84.5 AFRPD-53 608 49.0 55.3 97.0 57.4 94.8 59.1 97.6 65.7 97.3 39.6 87.5 28.3 85.5 AFRPD-53 1 024 23.0 56.4 98.1 60.4 95.4 59.0 97.7 66.3 96.9 40.6 87.4 29.5 85.2
Tools
Get Citation
Copy Citation Text
Saiqiang ZHANG, Shaofeng SI, Bin LU, Qing LI, Benyao CHEN, Anxin RONG. A Person Detection Algorithm in Fisheye Images Based on Rotated Boxes[J]. Acta Photonica Sinica, 2021, 50(10): 1010003
Paper Information
Category:
Received: Jun. 30, 2021
Accepted: Aug. 18, 2021
Published Online: Nov. 3, 2021
The Author Email: LI Qing (liqing@ime.ac.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20