Object detection method for aerial images based on adaptive slicing aided inference

Liwei JIN; Wangming XU; Yaoxiang LI

doi:10.37188/CJLCD.2024-0225

Chinese Journal of Liquid Crystals and Displays, Volume. 40, Issue 3, 472(2025)

Object detection method for aerial images based on adaptive slicing aided inference

Liwei JIN¹, Wangming XU^1,2、*, and Yaoxiang LI¹

Author Affiliations

¹School of Information Science and Engineering， Wuhan University of Science and Technology， Wuhan 430081， China

²Engineering Research Center for Metallurgical Automation and Detecting Technology of Ministry of Education， Wuhan University of Science and Technology， Wuhan 430081， China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. SAHI framework

Download full size

View in Article

Fig. 2. Workflow of the proposed method

Download full size

View in Article

Fig. 3. Diagram of score matrix structure

Download full size

View in Article

Fig. 4. Cropping diagram

Download full size

View in Article

Fig. 5. Examples of ATL

Download full size

View in Article

Fig. 6. Effect comparison for NMS

Download full size

View in Article

Fig. 7. Effect comparison for object detection on VisDrone dataset

Download full size

View in Article

Table 1. ASAI parameter optimization experiments

View table

View in Article

Table 1. ASAI parameter optimization experiments

Methods	mAP50/%	Time/ms
YOLOv8s （常规NMS）	30.1	54.2
+SAHI （常规NMS， k=16）	32.5	284.1
+ASAI （k=1）（常规NMS）	33.2	114.8
+ASAI （k=1）	33.4	110.5
+ASAI （k=2）	34.0	130.5
+ASAI （k=3）	34.7	145.9
+ASAI （k=4）	34.7	160.2
+ASAI （k=5）	34.8	187.2

Table 2. Comparison experiments of applying ASAI in typical lightweight object detection networks on VisDrone dataset

View table

View in Article

Table 2. Comparison experiments of applying ASAI in typical lightweight object detection networks on VisDrone dataset

Methods	mAP50/ %	Time/ ms	awning-tricycle/%	Bicycle/ %	Bus/ %	Car/ %	Motor/ %	Pedestrian/ %	People/ %	Tricycle/ %	Truck/ %	Van/ %
YOLOv7-tiny	27.9	55.1	5.2	9.2	36.2	67.8	32.4	32.3	27.4	13.5	25.4	29.2
YOLOv8n	14.8	51.0	1.3	1.5	24.4	58.5	7.8	13.9	41.3	4.7	13.4	18.4
YOLOv8s	30.1	54.2	8.3	10.5	46.9	70.9	30.0	33.1	16.7	21.5	29.6	33.4
YOLOv9-C	18.5	105.6	2.8	4.8	29.6	64.8	10.2	16.9	6.9	6.3	16.5	25.9
YOLOv7-tiny+ASAI	30.6	130.4	5.7	9.7	40.1	73.8	34.7	39.1	28.1	15.0	25.8	33.4
YOLOv8n+ASAI	19.5	99.4	1.9	2.8	34.1	68.2	13.7	22.3	7.6	6.4	16.4	21.8
YOLOv8s+ASAI	33.4	110.5	8.6	11.5	54.7	75.5	33.1	38.7	19.6	23.1	28.3	37.2
YOLOv9-C+ASAI	23.1	183.5	3.9	5.4	36.4	73.0	16.5	26.3	11.2	8.6	18.1	30.3

Table 3. Contrasts of YOLOv8s+ASAI combination method with the various new models

View table

View in Article

Table 3. Contrasts of YOLOv8s+ASAI combination method with the various new models

Methods	mAP50/%	mAP（50：95）/%	Parameters/M
Faster RCNN+ARFP^［13］	33.9	20.4	42.7+
UAV-YOLOv8s^［8］	47.0	29.2	10.3
Drone-YOLO（tiny）^［14］	42.8	25.6	5.4
MC-YOLOv5^［15］	45.5	26.6	38.2
YOLOv8s+ASAI	47.3	30.8	11.2

Table 4. Comparison experiments on AI-TOD dataset
View table
View in Article
Table 4. Comparison experiments on AI-TOD dataset
Methods mAP50/% Time/ms
YOLOv7-tiny 33.1 43.4
YOLOv8n 28.9 41.9
YOLOv8s 40.1 44.1
YOLOv9-C 46.2 53.5
YOLOv7-tiny+ASAI 34.1 83.7
YOLOv8n+ASAI 29.8 79.6
YOLOv8s+ASAI 41.3 76.1
YOLOv9-C+ASAI 47.2 97.6

Tools

Get Citation

Copy Citation Text

Liwei JIN, Wangming XU, Yaoxiang LI. Object detection method for aerial images based on adaptive slicing aided inference[J]. Chinese Journal of Liquid Crystals and Displays, 2025, 40(3): 472

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category:

Received: Aug. 4, 2024

Accepted: --

Published Online: Apr. 27, 2025

The Author Email: Wangming XU (xuwangming@wust.edu.cn)

DOI:10.37188/CJLCD.2024-0225

Topics

Table 1. ASAI parameter optimization experiments

Table 1. ASAI parameter optimization experiments

Table 2. Comparison experiments of applying ASAI in typical lightweight object detection networks on VisDrone dataset

Table 2. Comparison experiments of applying ASAI in typical lightweight object detection networks on VisDrone dataset

Table 3. Contrasts of YOLOv8s+ASAI combination method with the various new models

Table 3. Contrasts of YOLOv8s+ASAI combination method with the various new models

Table 4. Comparison experiments on AI-TOD dataset

Table 4. Comparison experiments on AI-TOD dataset