Image Segmentation and Computer Vision Datasets

This repository gathers a comprehensive collection of datasets used in Computer Vision and Image Segmentation.
It covers various domains such as semantic segmentation, instance segmentation, medical imaging, urban scenes, and interactive segmentation.

The goal is to provide a consolidated reference containing essential information — number of images, mask availability, resolution, dataset type, number of classes, description, and download links — to help researchers and developers choose suitable datasets for Deep Learning, Active Learning, Object Detection, and Scene Understanding tasks.

---

📊 Dataset Overview

The full list contains over 40 datasets. Click below to expand the table.

View the Full Dataset List (Click to expand)

Dataset Name	# Images	Masks	Size	Resolution	Kind of Dataset	# Classes	Description	Year	Link	Public?
VOC 2012	17,000	✅ Yes	4 GB	500×375	Object Segmentation	20	Includes training/validation/test splits with per-pixel annotations and object labels.	2012	Kaggle	✅
CityScapes	25,000	✅ Yes	25 GB	2048×1024	Urban Segmentation	30	50 different cities with pixel-level annotations for 30 classes.	2016	Official Site	✅
COCO	330,000	✅ Yes	50 GB	Variable	Object Segmentation	80	Complex scenes with multiple object masks.	2014	COCO	✅
LVIS	164,000	✅ Yes	25 GB	Variable	Instance Segmentation	1,203	Long-tail instance segmentation benchmark.	2019	LVIS	✅
ADE20K	27,000	✅ Yes	3 GB	Variable	Scene Parsing	150	Complete scene segmentation benchmark.	2016	MIT CSAIL	✅
GTA V Synthetic	25,000	✅ Yes	180 GB	1914×1052	Synthetic Semantic Segmentation	19	Synthetic urban scenes from GTA V with perfect pixel annotations.	2016	VISINF	✅
BraTS	3,000 (3D)	✅ Yes	200 GB	240×240×155	3D Medical Segmentation	3	Brain tumor dataset with edema, necrosis, and active tumor labels.	2012	CBICA	❌
LiTS	130 CT (3D)	✅ Yes	80 GB	512×512×Z	3D Medical Segmentation	2	3D liver and lesion segmentation dataset.	2017	CodaLab	❌
Kvasir-SEG	1,000	✅ Yes	2 GB	576×720	Medical Segmentation	1	Colorectal polyp dataset with binary masks.	2020	Simula	✅
Nuclei	30,000 patches	✅ Yes	100 MB	50×50	Biomedical Segmentation	1	Cell nuclei dataset with binary masks.	2018	Kaggle	✅
CVC-ClinicDB	612	✅ Yes	50 MB	384×288	Medical Segmentation	1	Colonoscopy frames for polyp detection.	2015	Kaggle	✅
REFUGE2	1,200	✅ Yes	3.8 GB	Variable	Medical Segmentation	2	Retinal disc and cup segmentation for glaucoma screening.	2020	Challenge	✅
ISIC	1,203,225	✅ Yes	Variable	Variable	Medical (Dermatology)	2–7	Massive dataset for skin lesion segmentation.	2016	ISIC Archive	✅
BrainMRI	3,929	✅ Yes	350 MB	256×256	Medical Segmentation	1	Brain tumor segmentation dataset.	2020	Kaggle	✅
LiverCT	131 CT (3D)	✅ Yes	80 GB	512×512×Z	3D Medical Segmentation	2	CT scans for liver injury segmentation.	2017	CodaLab	✅
RESC	110 scans	✅ Yes	500 MB	Variable	Medical Segmentation	3	Retinal edema segmentation dataset.	2018	GitHub	✅
TN3K	3,500	✅ Yes	200 MB	400×400	Medical Segmentation	1	Thyroid nodule ultrasound segmentation dataset.	2022	Kaggle	✅
DDTI	5,000	✅ Yes	1.5 GB	Variable	Medical Segmentation	1	Panoramic dental x-rays for teeth segmentation.	2022	Kaggle	✅
TG3K	3,100	✅ Yes	250 MB	400×400	Medical Segmentation	1	Ultrasound thyroid gland segmentation dataset.	2022	OpenMedLab	✅
BUSI	780	✅ Yes	250 MB	500×500	Medical Segmentation	3	Breast ultrasound segmentation dataset.	2019	Dataset Page	✅
CHAOS	80 scans (3D)	✅ Yes	20 GB	512×512×Z	3D Medical Segmentation	4	MRI and CT scans for liver, kidneys, and spleen segmentation.	2019	CHAOS	✅
ROCO	81,000	❌ No	8 GB	Variable	Medical Captioning	–	Radiology images paired with textual captions.	2018	GitHub	✅
MedPix	59,000	❌ No	Variable	Variable	Medical Image Database	–	Clinical and diagnostic image archive.	1999	MedPix	✅
NLPR	1,000 pairs	✅ Yes	998 MB	640×480	Salient Object Detection	1	Captured by Microsoft Kinect with indoor and outdoor scenes.	–	HyperAI	✅
PaviaU	1 image	❌ No	100 MB	610×340×103	Spectral Classification	9	Hyperspectral image captured over Pavia, Italy.	–	Kaggle	✅
BSDS500	500	✅ Yes	100 MB	Variable	Contour Detection	–	Human-annotated segmentation and contour detection benchmark.	–	Kaggle	✅
NYUV2	1,449	✅ Yes	5.5 GB	640×480	Indoor Scene Segmentation	40	RGB-D dataset captured using Microsoft Kinect.	2012	NYU	✅
SUNRGBD	10,335	✅ Yes	60 GB	Variable	2D/3D Segmentation	37	Densely annotated 3D indoor scenes.	2015	Princeton	✅
CamVid	701 frames	✅ Yes	570 MB	960×720	Video Semantic Segmentation	12	First video dataset with pixel-level annotations for urban scenes.	2008	CamVid	✅
300W-LP	122,450	❌ No	4 GB	Variable	Landmark Detection	68	Augmented version of 300W with rotated facial images.	2016	TensorFlow	✅
Visual Genome	108,000	❌ No	12 GB	Variable	Image Captioning	–	Object relationships and natural language annotations.	2016	VG	✅
ISPRS Vaihingen	33	✅ Yes	2 GB	~2500×2000	Aerial Image Segmentation	6	UHD aerial imagery with semantic labels.	2012	ISPRS	✅
NJU2K	1,985	✅ Yes	1.5 GB	Variable	Salient Object Detection	1	RGB image pairs for salient object detection.	2014	HyperAI	✅
STERE	1,000	✅ Yes	100 MB	1024×768	Object Detection	1	Stereo image pairs for object detection.	2015	KITTI	✅
GrabCut	50	✅ Yes	5 MB	Variable	Interactive Segmentation	1	Small dataset for interactive segmentation experiments.	2004	GitHub	✅
Awesome Medical Datasets	-	✅ Yes	-	-	Medical Image Segmentation	-	A collection of multiple open medical datasets.	-	OpenMedLab	✅
USPS	9,298	❌ No	10 MB	16×16	Classification	10	Handwritten digit dataset from postal codes.	1990	LibSVM	✅
MNIST	70,000	❌ No	15 MB	28×28	Classification	10	Classic handwritten digit dataset.	1998	Kaggle	✅
BioID	1,521	❌ No	150 MB	384×288	Face Detection	1	Grayscale face localization dataset.	1999	BioID	✅

---

🧩 Notes

• ✅ Public datasets are freely available for research and educational use.
• ❌ Non-public datasets may require registration, challenge participation, or access requests.
• Some datasets (e.g., LiTS, BraTS) are 3D volumetric and require preprocessing pipelines before use.

---

💡 How to Use

You can:

1. Explore datasets to benchmark segmentation models (e.g., U-Net, DeepLab, Mask R-CNN).
2. Use them in Active Learning or Continual Learning pipelines.
3. Combine multiple datasets to improve model generalization.

---

📚 Citation

If you use this list or parts of it, please cite this repository:

@misc{segmentation_datasets_collection,
  author = {Galetti, Daniel Martins},
  title = {Image Segmentation and Computer Vision Datasets Collection},
  year = {2025},
  url = {https://github.com/Danielgaletti/Datasets},
  note = {Comprehensive list of datasets for segmentation, detection, and scene understanding.}
}