Computer Vision Datasets Intermediate

Large-scale datasets that power modern computer vision research and applications. From the dataset that launched the deep learning revolution (ImageNet) to specialized datasets for autonomous driving and medical imaging.

ImageNet

14 million images, 1,000 classes (ILSVRC subset). The dataset that started the deep learning revolution. The ImageNet Large Scale Visual Recognition Challenge (ILSVRC) drove breakthroughs from AlexNet (2012) to modern Vision Transformers.

COCO (Common Objects in Context)

330,000 images with 2.5 million object instances. The standard benchmark for object detection, instance segmentation, keypoint detection, and image captioning.

from datasets import load_dataset

# Load COCO from Hugging Face
coco = load_dataset("detection-datasets/coco")

# Or with torchvision (requires downloaded images)
from torchvision.datasets import CocoDetection
coco = CocoDetection(root="coco/train2017", annFile="coco/annotations/instances_train2017.json")

Pascal VOC

11,530 images, 20 object classes. Predates COCO and remains a popular benchmark. Includes annotations for detection, segmentation, and action recognition.

ADE20K

25,210 images, 150 semantic categories. The standard benchmark for semantic segmentation. Dense pixel-level annotations covering indoor and outdoor scenes.

CelebA

202,599 face images of 10,177 celebrities. Each face annotated with 40 binary attributes (smiling, glasses, hat, etc.) and 5 landmark locations. Used for face attribute prediction, face generation, and recognition.

LFW (Labeled Faces in the Wild)

13,233 face images of 5,749 people. The standard benchmark for face verification. Given two face images, determine whether they are the same person.

Cityscapes

5,000 fine-annotated + 20,000 coarse images. Urban street scenes from 50 cities. Pixel-level semantic labels for 30 classes. Critical for autonomous driving research.

KITTI

Autonomous driving dataset. Includes stereo images, 3D point clouds (LiDAR), GPS, and annotated objects. Benchmarks for depth estimation, 3D detection, optical flow, and visual odometry.

Open Images

9 million images, 600 classes. Google's large-scale dataset with bounding boxes, segmentation masks, visual relationships, and localized narratives. One of the largest publicly available annotated datasets.

Comparison Table