Chinese Journal of Liquid Crystals and Displays, Volume. 38, Issue 11, 1468(2023)
Object detection algorithm based on adaptive focal CRIoU loss
Fig. 1. Difference of regression process of different boundary box loss functions
Fig. 4. Difference of loss in different regression states under inclusion
Fig. 6. Comparison of regression simulation convergence for different loss function
Fig. 7. Distribution of the last round regression error values of different loss function
Fig. 9. YOLOv3 convergence of different loss functions on PASCAL VOC2012
Fig. 10. Comparison of AF-CRIoU and NMS methods with different thresholds under AP75 condition
Fig. 11. Comparison of AF-CRIoU and NMS methods with different thresholds under AP50 condition
Table 1. Experimental hardware environment
View tableView in ArticleTable 1. Experimental hardware environment
实验硬件环境 环境配置 核心处理器 Intel Xeon Gold 5220@2.2 GHz 内存容量 256 GB 显卡型号 NVIDIA GeForce RTX 2080Ti
Table 2. Experimental software environment
View tableView in ArticleTable 2. Experimental software environment
实验软件环境 环境配置 服务器系统 Linux 7.6.1810 编程语言 Python 3.7.11 深度学习框架 Pytorch 1.7.1 开发工具 PyCharm 11.0.11；Matlab CUDA版本 10.1.243
Table 3. Comparison of YOLOv3 accuracy under different threshold
View tableView in ArticleTable 3. Comparison of YOLOv3 accuracy under different threshold
Loss function AP75/% AP50/% AF-CRIoU 65.45 59.1 SSE 60.31 54.7 Relative improv. 8.52 8.04 IoU 62.51 56.51 Relative improv. 4.7 4.58 GIoU 63.43 57.68 Relative improv. 3.18 2.46 CIoU 63.97 58.13 Relative improv. 2.31 1.67
Table 4. Comparison of Faster RCNN accuracy under diffe‑rent threshold
View tableView in ArticleTable 4. Comparison of Faster RCNN accuracy under diffe‑rent threshold
Loss function AP75/% AP50/% AF-CRIoU 56.61 52.28 Smooth L1 Loss 53.94 48.43 Relative improv. 4.94 7.9 IoU 54.62 48.78 Relative improv. 3.64 7.18 GIoU 54.93 48.93 Relative improv. 3.06 2.29 CIoU 55.21 50.91 Relative improv. 2.53 2.69
Table 5. Comparison of experimental results using different non maximum suppression methods and loss functions under AP75 conditions
View tableView in ArticleTable 5. Comparison of experimental results using different non maximum suppression methods and loss functions under AP75 conditions
NMS A-CRIoU-NMS AF-CRIoU AP75/% √ 65.09 √ 65.23 √ √ 65.45 √ √ 65.47
Table 6. Comparison of experimental results using different non maximum suppression methods and loss functions under AP50 conditions
View tableView in ArticleTable 6. Comparison of experimental results using different non maximum suppression methods and loss functions under AP50 conditions
NMS A-CRIoU-NMS AF-CRIoU AP50/% √ 58.52 √ 58.90 √ √ 59.10 √ √ 59.13
Table 7. Comparison of detection accuracy
View tableView in ArticleTable 7. Comparison of detection accuracy
Loss hat/% nohat/% mAP/% IoU 68.31 65.72 67.01 GIoU 70.47 69.87 70.17 CIoU 72.92 71.38 72.15 AF-CRIoU 74.28 72.57 73.43
Get Citation
Copy Citation Text
Zhen-jiu XIAO, Hao-ze ZHAO, Li-li ZHANG, Yu XIA, Jie-long GUO, Hui YU, Cheng-long LI, Li-wen WANG. Object detection algorithm based on adaptive focal CRIoU loss[J]. Chinese Journal of Liquid Crystals and Displays, 2023, 38(11): 1468
Category: Research Articles
Received: Jan. 6, 2023
Accepted: --
Published Online: Nov. 29, 2023
The Author Email: Jie-long GUO (gjl@fjirsm.ac.cn)
