计算机视觉常见数据集

ImageNet

overview

(updated on April 30, 2010)

• Total number of non-empty synsets: 21841
• Total number of images: 14,197,122
• Number of images with bounding box annotations: 1,034,908
• Number of synsets with SIFT features: 1000

• Number of images with SIFT features: 1.2 million

• 总类别: 21,841

• 总图像: 14,197,122
• 有bounding box标注的图像数量: 1,034,908
• 有SIFT特征的类别数: 1000
• 有SIFT特征的图像数: 1.2 million

tasks

• Scene classification/场景分类
• Object localization/目标定位
• Object detection/目标检测
• Object detection from video/视频目标检测

---

MSCOCO

overview

• Object segmentation
• Recognition in Context
• Multiple objects per image
• More than 300,000 images
• More than 2 Million instances
• 80 object categories
• 5 captions per image

• Keypoints on 100,000 people

• 目标分割

• 内容识别
• 每幅图像多个实例
• 图像:300,000+
• 实例:2,000,000+
• 目标种类:80
• 平均每张图像标注数:5
• 有关键点人数:100,000

tasks

• Detection/检测
• Keypoints/关键点
• Captioning/标注

---

Open Images

• images URL: ~9,000,000
• 2,000,000 bounding boxes spanning 600 object classes (1.24M in train, 830K in validation+test)
• 4,300,000 human-verified positive image-level labels on the training set
• coming soon: Trained models (both image-level and object detectors).

---

Youtube-8M

(2017 update)
- Video URLs: 7,000,000
- Video: 450,000 hours
- Audio/Visual Features: 3,200,000,000
- Classes: 4716
- Avg.Labels/Video: 3.4

悲哀的是大天朝用不了youtube

---

SUN

overview

• Images: 131,067
• Scene categories: 908
• Segmented objects: 313,884
• Object categories: 4,479
• 图像数: 131,067
• 场景种类: 908
• 分割种类: 313,884
• 物体种类: 4,479

tasks

• Scene Recognition Benchmark（场景识别）
• scene categories: 397
• Object Detection Benchmark（目标检测）
• images: 16,873

---

NUS-WIDE

• images: 269,648
• unique tags: 5,018
• low-level features types: 6

---

PASCAL VOC 2010

• Training and validation images: 10,103
• Testing images: 9,637
• categories: 33/59