Group-Mix SAM: Lightweight Segmentation Solution for Industrial Implementations

Since its emergence in early 2023, Segment Anything Model(SAM) has garnered significant academic interests and spawned many investigations from various perspectives due to its zero-shot generalization ability. However, the deployment of SAM in scenes that required real-time performance, such as indu...

Full description

Saved in:
Bibliographic Details
Published in2024 IEEE International Conference on Cognitive Computing and Complex Data (ICCD) pp. 18 - 22
Main Authors Wu, Liang, Duan, Puhong, Ma, Xiaoguang
Format Conference Proceeding
LanguageEnglish
Published IEEE 28.09.2024
Subjects
Assembly
Cognitive systems
Computational modeling
Computer architecture
Data models
Image segmentation
Industrial implementations
Knowledge distillation
Real-time systems
Segment anything model
Online AccessGet full text

Cover

More Information
Summary:Since its emergence in early 2023, Segment Anything Model(SAM) has garnered significant academic interests and spawned many investigations from various perspectives due to its zero-shot generalization ability. However, the deployment of SAM in scenes that required real-time performance, such as industrial assembly liners, had severe challenges due to its heavy image encoder, whose parameters were as large as 632M. In this study, we proposed a lightweight segmentation solution for industrial assembly line applications, wherein the heavyweight image encoder was replaced with a lightweight one. Specifically, we employed decoupled distillation to train the encoder of MobileSAM in a resource-limited setting. The entire knowledge distillation experiment could be completed in a single day on a single RTX 4090. The resulting lightweight SAM, called Group- Mix SAM, had 37.63% (2.16M) fewer parameters and 42.5% (15614.7M) fewer floating-point operations compared to MobileSAM. Additionally, on our homemade industrial dataset: MALD, its mIoU was only marginally lower than that of MobileSAM, at 0.615. Finally, we conducted comprehensive experiments to demonstrate the superiority of Group-Mix SAM in real industrial scenes.
DOI:10.1109/ICCD62811.2024.10843550