Multi-Head Mixed Self-Attention Mechanism for Object Detection

Table 1. Experimental results on BDD100k dataset for different numbers of heads within MMSA
View table
Table 1. Experimental results on BDD100k dataset for different numbers of heads within MMSA
Number of head mAP@0.5 /% Model size /MB
4 52.2 6.14
8 52.4 6.14
16 52.5 6.14
32 52.2 6.14

Table 2. Experimental results on BDD100k dataset for different sizes of local pooling region within MMSA
View table
Table 2. Experimental results on BDD100k dataset for different sizes of local pooling region within MMSA
Size of local pooling region mAP@0.5 /% Model size /MB
1 52.3 6.14
2 52.6 6.14
5 52.5 6.14
10 51.8 6.14

Table 3. Comparison results of adding different MMSA on different parts tested in BDD100K dataset
View table
Table 3. Comparison results of adding different MMSA on different parts tested in BDD100K dataset
Number of head Size of local pooling region Backbone Neck mAP@0.5 /% Model size /MB
16 2 √ × 52.7 6.16
16 2 × √ 52.6 6.14
16 2 √ √ 52.5 6.17
8 5 √ × 52.8 6.16
8 5 × √ 52.4 6.14
8 5 √ √ 52.5 6.17

Table 4. Comparison results of adding different MMSA on different parts tested in TinyPerson dataset
View table
Table 4. Comparison results of adding different MMSA on different parts tested in TinyPerson dataset
Number of head Size of local pooling region Backbone Neck mAP@0.5 /% Model size /MB
16 2 √ × 15.5 6.18
16 2 × √ 14.9 6.16
16 2 √ √ 16.0 6.19
8 5 √ × 16.0 6.18
8 5 × √ 15.9 6.16
8 5 √ √ 16.1 6.19

Table 5. mAP@0.5 of using GAP and LMP in MMSA
View table
Table 5. mAP@0.5 of using GAP and LMP in MMSA
Algorithm CityPersons CrowdHuman TT100K BDD100K TinyPerson
YOLOv8n 61.3 69.2 74.6 51.8 14.3
MMSA-GAP 61.2 69.6 75.2 52.5 15.6
MMSA-LMP 62.2 70.1 76.9 52.8 16.0

Table 6. mAP@0.5 of different models on various datasets
View table
Table 6. mAP@0.5 of different models on various datasets
Algorithm CityPersons CrowdHuman TT100K BDD100K TinyPerson
YOLOv5n 58.0 65.2 55.1 47.6 12.5
YOLOv8n 61.3 69.2 74.6 51.8 14.3
YOLOv10n 58.7 68.7 72.7 50.6 13.7
YOLOv8n+MMSA 62.2 70.1 76.9 52.8 16.0

Table 7. mAP@0.5 of different datasets for incorporating various attention mechanisms

View table

Table 7. mAP@0.5 of different datasets for incorporating various attention mechanisms

Algorithm	CityPersons	CrowdHuman	TT100K	BDD100K	TinyPerson
YOLOv8n	61.3	69.2	74.6	51.8	14.3
YOLOv8n+CBAM	59.6	68.2	73.6	49.8	15.1
YOLOv8n+SENet	59.7	68.4	71.8	50.2	14.8
YOLOv8n+CA	59.7	68.7	73.6	49.7	15.2
YOLOv8n+MMSA	62.2	70.1	76.9	52.8	16.0

Table 8. Comparison of model size and detection speed before and after adding MMSA
View table
Table 8. Comparison of model size and detection speed before and after adding MMSA
Model FPS /（frame·s^-1） Memory /MB GFLOPS
YOLOv8n 184 5.97 8.1
YOLOv8n+MMSA 145 6.18 8.1

Tools

Get Citation

Copy Citation Text

Qinghua Su, Jianhong Mu, Wenhui Liang, Xiyu Wang, Juntao Li. Multi-Head Mixed Self-Attention Mechanism for Object Detection[J]. Laser & Optoelectronics Progress, 2025, 62(6): 0637006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Digital Image Processing

Received: Jun. 19, 2024

Accepted: Aug. 1, 2024

Published Online: Mar. 6, 2025

The Author Email:

DOI:10.3788/LOP241509

CSTR:32186.14.LOP241509

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Experimental results on BDD100k dataset for different numbers of heads within MMSA

Table 1. Experimental results on BDD100k dataset for different numbers of heads within MMSA

Table 2. Experimental results on BDD100k dataset for different sizes of local pooling region within MMSA

Table 2. Experimental results on BDD100k dataset for different sizes of local pooling region within MMSA

Table 3. Comparison results of adding different MMSA on different parts tested in BDD100K dataset

Table 3. Comparison results of adding different MMSA on different parts tested in BDD100K dataset

Table 4. Comparison results of adding different MMSA on different parts tested in TinyPerson dataset

Table 4. Comparison results of adding different MMSA on different parts tested in TinyPerson dataset

Table 5. mAP@0.5 of using GAP and LMP in MMSA

Table 5. mAP@0.5 of using GAP and LMP in MMSA

Table 6. mAP@0.5 of different models on various datasets

Table 6. mAP@0.5 of different models on various datasets

Table 7. mAP@0.5 of different datasets for incorporating various attention mechanisms

Table 7. mAP@0.5 of different datasets for incorporating various attention mechanisms

Table 8. Comparison of model size and detection speed before and after adding MMSA

Table 8. Comparison of model size and detection speed before and after adding MMSA