Lightweight Small Object Detection Algorithm Based on STD-DETR

Fig. 8. Comparison of detection results on the VisDrone2019 dataset before and after model improvement. (a) Dense targets at daytime; (b) sparse targets at daytime; (c) dense targets at night; (d) sparse targets at night

Download full size

Fig. 9. Confusion matrixes before and after model improvement. (a) RT-DETR-R18 model; (b) STD-DETR model

Download full size

Fig. 10. Comparison of detection results before and after model improvement. (a) Land targets; (b) sea targets

Download full size

Table 1. Ablation experimental results on the VisDrone2019 dataset

View table

Table 1. Ablation experimental results on the VisDrone2019 dataset

RT-DETR-R18	Starnet	STOS	PIoU	P /%	R /%	mAP₅₀ /%	mAP_50∶95 /%	Param /M	GFLOPs
√				62.8	46.5	47.7	28.9	19.8	57.3
√	√			62.5	46.1	47.3	28.6	12.1	32.6
√		√		64.5	47.7	49.3	30.2	20.0	63.2
√			√	64.3	47.5	49.1	30.0	19.7	57.0
√	√	√		63.2	47.7	48.7	29.6	12.3	37.0
√	√		√	63.3	47.0	48.2	29.0	12.1	33.0
√		√	√	64.4	48.0	49.8	30.5	20.0	66.6
√	√	√	√	64.1	48.7	50.0	30.4	12.3	37.8

Table 2. Performance comparison of different loss functions
View table
Table 2. Performance comparison of different loss functions
Loss function P R mAP₅₀ mAP_50∶95
GIoU 62.8 46.5 47.7 28.9
DIoU 63.7 47.1 48.6 29.6
SIoU 63.3 47.0 48.5 29.6
ShapeIoU 63.1 46.8 48.4 29.4
Focaler-ShapeIoU 64.5 46.9 49.0 29.7
Focaler-PIoU 63.5 46.6 48.6 29.6
PIoU 64.3 47.5 49.1 30.0

Table 3. Comparison of experimental results between STOS and P2 models
View table
Table 3. Comparison of experimental results between STOS and P2 models
Model
P /
%
R /
%
mAP₅₀ /
%
mAP_50∶95 /
%
Param /M GFLOPs
STOS 64.5 47.7 49.3 30.2 20.0 63.2
P2 63.4 47.8 49.5 30.6 19.6 81.7

Table 4. Detection results by different models on the VisDrone2019 dataset

View table

Table 4. Detection results by different models on the VisDrone2019 dataset

Model	P /%	R /%	mAP₅₀ /%	mAP_50∶95 /%	Param /M	GFLOPs
Faster-R-CNN	34.6	36.8	30.9	13.1	63.2	370.0
SSD	21.1	35.8	24.0	11.9	12.3	63.2
YOLOv5m	50.3	37.9	36.3	19.2	21.2	48.3
YOLOv5l	45.1	35.2	38.7	24.3	45.9	108.4
YOLOXs	35.2	41.4	34.3	20.0	9.0	26.8
YOLOv7	54.1	43.6	42.8	22.5	37.2	103.3
YOLOv8m	55.7	44.3	40.9	24.3	25.8	78.7
YOLOv8l	57.4	45.3	45.7	28.1	43.6	165.2
YOLOv7+Byter	57.1	46.5	45.8	26.3	34.2	—
BiEO-YOLOv8s	56.0	44.0	46.8	—	11.3	—
RT-DETR-R18	62.8	46.5	47.7	28.9	19.8	57.3
STD-DETR	64.1	48.7	50.0	30.4	12.3	37.8

Table 5. Comparison of detection results on the Tinyperson dataset
View table
Table 5. Comparison of detection results on the Tinyperson dataset
Model P /% R /% mAP₅₀ /% mAP_50∶95 /% Param /M GFLOPs
SC-YOLO — — 15.7 4.7 12.1 —
RT-DETR-R18 33.5 20.2 15.4 5.2 19.8 57.3
STD-DETR 37.2 22.1 18.5 6.1 12.2 37.1

Tools

Get Citation

Copy Citation Text

Zeyu Yin, Bo Yang, Jinling Chen, Chuangchuang Zhu, Hongli Chen, Jin Tao. Lightweight Small Object Detection Algorithm Based on STD-DETR[J]. Laser & Optoelectronics Progress, 2025, 62(8): 0815002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites