ResNeSt: Split-Attention Networks
Hang Zhang1 Chongruo Wu2 Zhongyue Zhang3 Yi Zhu4 Haibin Lin5 Zhi Zhang4 Yue Sun6 Tong He4 Jonas Mueller4 R. Manmatha4 Mu Li4 Alexander Smola4 Facebook1 UC Davis2 Snap3 Amazon4 ByteDance5 SenseTime6
Abstract
It is well known that featuremap attention and multi-path representation are important for visual recognition. In this paper, we present a modularized architecture, which applies the channel-wise attention on different network branches to leverage their success in capturing cross-feature interactions and learning diverse representations. Our design results in a simple and unified computation block, which can be parameterized using only a few variables. Our model, named ResNeSt, outperforms EfficientNet in accuracy and latency trade-off on image classification. In addition, ResNeSt has achieved superior transfer learning results on several public benchmarks serving as the backbone, and has been adopted by the winning entries of COCO-LVIS challenge. The source code for complete system and pretrained models are publicly available.
中文速览
视觉识别领域长期面临如何同时利用"通道注意力"和"多路径特征表示"这两种优势的难题。研究者提出了一种名为 ResNeSt 的新型网络模块——Split-Attention Block（分裂注意力模块），其核心思路是把输入特征图拆分成若干组，在每个组内通过多个子分支提取不同表示，再用通道级软注意力机制对各分支的结果进行加权融合，从而让网络既能捕捉跨通道的特征相关性，又能保留多样化的独立表示。整个模块仅用"基数 K"和"径向数 R"两个超参数即可完整描述，并可用标准卷积算子高效实现。在 ImageNet 图像分类上，ResNeSt 的精度-延迟权衡优于经神经架构搜索得到的 EfficientNet；在目标检测、实例分割和语义分割等迁移学习任务上也创造了多个榜单最优成绩，并被 2020 年 COCO-LVIS 竞赛冠军方案采用。这项工作的价值在于，它用一个简洁统一的模块设计将注意力机制与多路径结构有机结合，为后续骨干网络设计和神经架构搜索提供了新的高效构建单元。
原文 arXiv:2004.08955;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2004.08955v2