Microsoft COCO：上下文中的常见物体

Microsoft COCO: Common Objects in Context

林宗毅 Tsung-Yi Lin · Microsoft · 2014-05-01 · arXiv:1405.0312 ↗ · 被引 53675

摘要 · Abstract

我们提出了一个新数据集，旨在通过将物体识别问题置于场景理解这一更广泛问题的背景下，推动物体识别技术的发展。为此，我们收集了包含常见物体在其自然环境中出现的复杂日常场景图像。物体使用逐实例分割进行标注，以帮助精确定位物体。我们的数据集包含 91 种物体类型的照片，这些类型对于 4 岁儿童来说易于识别。数据集共有 328k 张图像中的 250 万个标注实例，其创建过程通过新颖的用户界面（用于类别检测、实例定位和实例分割）吸引了大量众包工人的参与。我们提供了与 PASCAL、ImageNet 和 SUN 数据集的详细统计分析比较。最后，我们使用可变形部件模型提供了边界框和分割检测结果的基线性能分析。

We present a new dataset with the goal of advancing the state-of-the-art in object recognition by placing the question of object recognition in the context of the broader question of scene understanding. This is achieved by gathering images of complex everyday scenes containing common objects in their natural context. Objects are labeled using per-instance segmentations to aid in precise object localization. Our dataset contains photos of 91 objects types that would be easily recognizable by a 4 year old. With a total of 2.5 million labeled instances in 328k images, the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation. We present a detailed statistical analysis of the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide baseline performance analysis for bounding box and segmentation detection results using a Deformable Parts Model.

核心贡献 · Key contributions

提出 MS COCO 数据集，包含 91 个物体类别、328k 张图像中的 250 万个标注实例。
Introduces MS COCO dataset with 91 object categories, 2.5M instances in 328k images.
聚焦非典型视角图像，强调上下文关系与精确的实例分割。
Focuses on non-iconic images with contextual relationships and precise instance segmentation.
创新标注流程，采用分层标注与众包提高效率。
Novel annotation pipeline using hierarchical labeling and crowdsourcing for efficiency.
与 PASCAL、ImageNet 和 SUN 数据集进行详细统计对比。
Provides detailed statistical comparison with PASCAL, ImageNet, and SUN datasets.
使用可变形部件模型提供检测与分割的基线结果。
Baseline detection and segmentation results using Deformable Parts Model.
展示数据集难度及跨数据集泛化的优势。
Demonstrates dataset difficulty and cross-dataset generalization benefits.

局限 · Limitations

仅包含“物体”类别，不包括“材质”如天空或草地。
Only includes 'thing' categories, not 'stuff' like sky or grass.
限于 91 个基础类别，不包括细粒度类别。
Limited to 91 entry-level categories; fine-grained categories excluded.
尽管有验证，众包标注仍可能存在质量差异。
Crowd-sourced annotations may have quality variance despite verification.
基线模型（DPM）可能无法完全捕捉数据集的复杂性。
Baseline models (DPM) may not capture full complexity of dataset.
数据集分为 2014 和 2015 版本，部分类别未分割。
Dataset split into 2014 and 2015 releases; some categories not segmented.

论文章节 · Sections（共 19）

摘要 Abstract
1 引言 1 Introduction
2 相关工作 2 Related Work
3 图像收集 3 Image Collection
3.1 常见物体类别 3.1 Common Object Categories
3.2 非标志性图像收集 3.2 Non-iconic Image Collection
4 图像标注 4 Image Annotation
4.1 类别标注 4.1 Category Labeling
4.2 实例检测 4.2 Instance Spotting
4.3 实例分割 4.3 Instance Segmentation
4.4 标注性能分析 4.4 Annotation Performance Analysis
4.5 描述标注 4.5 Caption Annotation
5 数据集统计 5 Dataset Statistics
6 数据集划分 6 Dataset Splits
7 算法分析 7 Algorithmic Analysis
8 讨论 8 Discussion
附录概述 Appendix Overview
附录 I：用户界面 Appendix I: User Interfaces
附录 II：物体与场景类别 Appendix II: Object & Scene Categories

阅读逐段中英对照全文 →