ImageNet 大规模视觉识别挑战赛

ImageNet Large Scale Visual Recognition Challenge

李飞飞 Fei-Fei Li · Stanford · 2014-09-01 · arXiv:1409.0575 ↗ · 被引 43007

打开互动全文版（逐段中英对照 + 图/公式 + 论文问答）→

摘要 · Abstract

ImageNet 大规模视觉识别挑战赛是一个针对数百个物体类别和数百万张图像的物体分类与检测基准。该挑战赛自 2010 年起每年举办，吸引了超过五十家机构参与。本文描述了该基准数据集的创建过程以及由此带来的物体识别进展。我们讨论了收集大规模真实标注的挑战，强调了分类物体识别的关键突破，详细分析了大规模图像分类和物体检测领域的现状，并将最先进的计算机视觉精度与人类精度进行了比较。最后，我们总结了挑战赛五年来的经验教训，并提出了未来的方向和改进措施。

The ImageNet Large Scale Visual Recognition Challenge is a benchmark in object category classification and detection on hundreds of object categories and millions of images. The challenge has been run annually from 2010 to present, attracting participation from more than fifty institutions. This paper describes the creation of this benchmark dataset and the advances in object recognition that have been possible as a result. We discuss the challenges of collecting large-scale ground truth annotation, highlight key breakthroughs in categorical object recognition, provide a detailed analysis of the current state of the field of large-scale image classification and object detection, and compare the state-of-the-art computer vision accuracy with human accuracy. We conclude with lessons learned in the five years of the challenge, and propose future directions and improvements.

核心贡献 · Key contributions

• 提出了 ImageNet 大规模视觉识别挑战赛（ILSVRC），这是一个包含 1000 个类别和数百万张图像的物体分类与检测基准。
  Introduced the ImageNet Large Scale Visual Recognition Challenge (ILSVRC), a benchmark for object classification and detection with 1000 categories and millions of images.
• 开发了可扩展的众包方法用于大规模图像标注，达到了 99.7%的精确率。
  Developed scalable crowdsourcing methods for large-scale image annotation, achieving 99.7% precision.
• 证明了深度卷积神经网络（如 AlexNet）显著优于传统方法，标志着物体识别的转折点。
  Demonstrated that deep convolutional neural networks (e.g., AlexNet) significantly outperform traditional methods, marking a turning point in object recognition.
• 提供了五年间算法进展的详细分析，将最先进的计算机视觉准确率与人类准确率进行了比较。
  Provided a detailed analysis of algorithmic progress over five years, comparing state-of-the-art computer vision accuracy with human accuracy.
• 建立了大规模分类、定位和检测的评估指标，包括 top-5 错误率和平均精确率。
  Established evaluation metrics for large-scale classification, localization, and detection, including top-5 error and average precision.
• 发布了公开可用的数据集和竞赛框架，成为物体识别研究的标准基准。
  Released a publicly available dataset and competition framework that became the standard benchmark for object recognition research.

局限 · Limitations

• 数据集标注仅对每张图像标记一个物体类别用于分类，当存在多个物体时会产生歧义。
  Dataset annotation only labels one object category per image for classification, causing ambiguity when multiple objects are present.
• 定位评估指标对于包含聚集物体的图像可能不准确，需要手动移除 3.5%的图像。
  Evaluation metrics for localization may be inaccurate for images with clustered objects, requiring manual removal of 3.5% of images.
• 小物体实例对定位评估构成挑战；需要对小于 25x25 像素的物体调整阈值。
  Small object instances pose challenges for localization evaluation; threshold adjustment is needed for objects under 25x25 pixels.
• 数据集的 1000 个类别仅限于 ImageNet 同义词集，可能无法覆盖所有真实世界的物体类型。
  The dataset's 1000 categories are limited to ImageNet synsets and may not cover all real-world object types.
• 众包标注质量依赖于工人共识，不同物体类别之间可能存在差异。
  Crowdsourcing annotation quality relies on worker consensus, which may vary across different object categories.

论文章节 · Sections（共 33）

• 摘要 Abstract
• 1 引言 1 Introduction
• 1.1 相关工作 1.1 Related work
• 1.2 论文结构 1.2 Paper layout
• 2 挑战任务 2 Challenge tasks
• 2.1 图像分类任务 2.1 Image classification task
• 2.2 单目标定位任务 2.2 Single-object localization task
• 2.3 目标检测任务 2.3 Object detection task
• 3 大规模数据集构建 3 Dataset construction at large scale
• 3.1 图像分类数据集构建 3.1 Image classification dataset construction
• 3.2 单目标定位数据集构建 3.2 Single-object localization dataset construction
• 3.3 目标检测数据集构建 3.3 Object detection dataset construction
• 4 大规模评估 4 Evaluation at large scale
• 4.1 图像分类 4.1 Image classification
• 4.2 单目标定位 4.2 Single-object localization
• 4.3 目标检测 4.3 Object detection
• 5 方法 5 Methods
• 5.1 挑战赛参赛作品 5.1 Challenge entries
• 5.2 大规模算法创新 5.2 Large scale algorithmic innovations
• 6.1 多年来的改进 6.1 Improvements over the years
• 6.2 统计显著性 6.2 Statistical significance
• 6.3 类别物体识别的当前状态 6.3 Current state of categorical object recognition
• 6.4 大规模图像分类中的人类准确率 6.4 Human accuracy on large-scale image classification
• 7 结论 7 Conclusions
• 7.1 经验教训 7.1 Lessons learned
• 7.2 批评 7.2 Criticism
• 7.3 未来 7.3 The future
• 附录 A ILSVRC2012-2014 图像分类与单目标定位类别 Appendix A ILSVRC2012-2014 image classification and single-object localization object categories
• 附录 B 单目标定位数据集补充统计 Appendix B Additional single-object localization dataset statistics
• 附录 C 手动筛选的物体检测场景图像查询 Appendix C Manually curated queries for obtaining object detection scene images
• 附录 D 全图标注问题层级结构 Appendix D Hierarchy of questions for full image annotation
• 附录 E 物体检测边界框系统修改 Appendix E Modification to bounding box system for object detection
• 附录 F 竞赛协议 Appendix F Competition protocol

阅读逐段中英对照全文 →