首页|Studies from Nanjing University of Aeronautics and Astronautics in the Area of C omputational Intelligence Described (Muster: a Multi-scale Transformer-based Dec oder for Semantic Segmentation)
Studies from Nanjing University of Aeronautics and Astronautics in the Area of C omputational Intelligence Described (Muster: a Multi-scale Transformer-based Dec oder for Semantic Segmentation)
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
NSTL
2024 OCT 03 (NewsRx)-By a News Reporter-Staff News Editor at Robotics & Machine Learning Daily News Daily News-Data detailed on Machine Learning - Co mputational Intelligence have been presented. According to news reporting out of Nanjing, People's Republic of China, by NewsRx editors, research stated, "In re cent works on semantic segmentation, there has been a significant focus on desig ning and integrating transformer-based encoders. However, less attention has bee n given to transformer-based decoders." Financial support for this research came from National Natural Science Foundatio n of China (NSFC). Our news journalists obtained a quote from the research from the Nanjing Univers ity of Aeronautics and Astronautics, "We emphasize that the decoder stage is equ ally vital as the encoder in achieving superior segmentation performance. It dis entangles and refines high-level cues, enabling precise object boundary delineat ion at the pixel level. In this paper, we introduce a novel transformer-based de coder called MUSTER, which seamlessly integrates with hierarchical encoders and consistently delivers high-quality segmentation results, regardless of the encod er architecture. Furthermore, we present a variant of MUSTER that reduces FLOPS while maintaining performance. MUSTER incorporates carefully designed multi-head skip attention (MSKA) units and introduces innovative upsampling operations. Th e MSKA units enable the fusion of multi-scale features from the encoder and deco der, facilitating comprehensive information integration. The upsampling operatio n leverages encoder features to enhance object localization and surpasses tradit ional upsampling methods, improving mIoU (mean Intersection over Union) by 0.4% to 3.2%. On the challenging ADE20K dataset, our best model achieves a single-scale mIoU of 50.23 and a multi-scale mIoU of 51.88, which is on-par w ith the current state-of-the-art model."
NanjingPeople's Republic of ChinaAsi aComputational IntelligenceMachine LearningNanjing University of Aeronauti cs and Astronautics
NETL
NSTL
