Object Detection Datasets Overview#

Training a robust and accurate object detection model requires a comprehensive dataset. This guide introduces various formats of datasets that are compatible with the Ultralytics YOLO model and provides insights into their structure, usage, and how to convert between different formats.

Supported Dataset Formats#

Ultralytics YOLO format#

The Ultralytics YOLO format is a dataset configuration format that allows you to define the dataset root directory, the relative paths to training/validation/testing image directories or *.txt files containing image paths, and a dictionary of class names. Here is an example:

# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license

# COCO8 dataset (first 8 images from COCO train2017) by Ultralytics
# Documentation: https://docs.ultralytics.com/datasets/detect/coco8
# Example usage: yolo train data=coco8.yaml
# parent
# ├── ultralytics
# └── datasets
#     └── coco8 ← downloads here (1 MB)

# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]
path: coco8 # dataset root dir
train: images/train # train images (relative to 'path') 4 images
val: images/val # val images (relative to 'path') 4 images
test: # test images (optional)

# Classes
names:
  0: person
  1: bicycle
  2: car
  3: motorcycle
  4: airplane
  5: bus
  6: train
  7: truck
  8: boat
  9: traffic light
  10: fire hydrant
  11: stop sign
  12: parking meter
  13: bench
  14: bird
  15: cat
  16: dog
  17: horse
  18: sheep
  19: cow
  20: elephant
  21: bear
  22: zebra
  23: giraffe
  24: backpack
  25: umbrella
  26: handbag
  27: tie
  28: suitcase
  29: frisbee
  30: skis
  31: snowboard
  32: sports ball
  33: kite
  34: baseball bat
  35: baseball glove
  36: skateboard
  37: surfboard
  38: tennis racket
  39: bottle
  40: wine glass
  41: cup
  42: fork
  43: knife
  44: spoon
  45: bowl
  46: banana
  47: apple
  48: sandwich
  49: orange
  50: broccoli
  51: carrot
  52: hot dog
  53: pizza
  54: donut
  55: cake
  56: chair
  57: couch
  58: potted plant
  59: bed
  60: dining table
  61: toilet
  62: tv
  63: laptop
  64: mouse
  65: remote
  66: keyboard
  67: cell phone
  68: microwave
  69: oven
  70: toaster
  71: sink
  72: refrigerator
  73: book
  74: clock
  75: vase
  76: scissors
  77: teddy bear
  78: hair drier
  79: toothbrush

# Download script/URL (optional)
download: https://github.com/ultralytics/assets/releases/download/v0.0.0/coco8.zip

Each of train, val, and test accepts a directory, a list of directories, or a *.txt file listing one image path per line (paths starting with ./ resolve relative to the *.txt file). A *.txt file is useful to train on a subset of a directory, skip unlabeled images, or combine images from multiple sources into one split.

path: datasets/coco8 # dataset root
train: train.txt # a directory, a list e.g. [images/a, images/b], or a *.txt file
val: val.txt
names:
  0: person

Labels for this format should be exported to YOLO format with one *.txt file per image. If there are no objects in an image, no *.txt file is required. The *.txt file should be formatted with one row per object in class x_center y_center width height format. Box coordinates must be in normalized xywh format (from 0 to 1). If your boxes are in pixels, you should divide x_center and width by image width, and y_center and height by image height. Class numbers should be zero-indexed (start with 0).

The label file corresponding to the above image contains 2 persons (class 0) and a tie (class 27):

When using the Ultralytics YOLO format, organize your training and validation images and labels as shown in the COCO8 dataset example below.

Usage Example#

Here's how you can use YOLO format datasets to train your model:

from ultralytics import YOLO

# Load a model
model = YOLO("yolo26n.pt")  # load a pretrained model (recommended for training)

# Train the model
results = model.train(data="coco8.yaml", epochs=100, imgsz=640)

Ultralytics NDJSON format#

The NDJSON (Newline Delimited JSON) format provides an alternative way to define datasets for Ultralytics YOLO models. This format stores dataset metadata and annotations in a single file where each line contains a separate JSON object.

An NDJSON dataset file contains:

1. Dataset record (first line): Contains dataset metadata including task type, class names, and general information
2. Image records (subsequent lines): Contains individual image data including dimensions, annotations, and file paths

{
    "type": "dataset",
    "task": "detect",
    "name": "Example",
    "description": "COCO NDJSON example dataset",
    "url": "https://app.ultralytics.com/user/datasets/example",
    "class_names": { "0": "person", "1": "bicycle", "2": "car" },
    "bytes": 426342,
    "version": 0,
    "created_at": "2024-01-01T00:00:00Z",
    "updated_at": "2025-01-01T00:00:00Z"
}

Custom image metadata#

Each image record may include a metadata JSON object for application-specific context such as capture conditions, equipment identifiers, or review status. Nested values are supported. When imported into Ultralytics Platform, the metadata is stored with that image and can be viewed or edited from its fullscreen information panel.

{
    "type": "image",
    "file": "airbus-wing.jpg",
    "url": "https://example.com/airbus-wing.jpg",
    "split": "train",
    "metadata": { "aircraft": { "family": "A350", "section": "wing" }, "inspectionStatus": "reviewed" }
}

Platform limits top-level metadata keys to 128 characters, each image's serialized metadata object to 500,000 characters, and the combined effective metadata in one NDJSON import to 500,000 characters.

Usage Example#

To use an NDJSON dataset with YOLO26, simply specify the path to the .ndjson file:

from ultralytics import YOLO

# Load a model
model = YOLO("yolo26n.pt")

# Train using NDJSON dataset
results = model.train(data="path/to/dataset.ndjson", epochs=100, imgsz=640)

Advantages of NDJSON format#

• Single file: All dataset information contained in one file
• Streaming: Can process large datasets line-by-line without loading everything into memory
• Cloud integration: Supports remote image URLs for cloud-based training
• Extensible: Easy to add custom metadata fields
• Version control: Single file format works well with git and version control systems

Supported Datasets#

Here is a list of the supported datasets and a brief description for each:

• African-wildlife: A dataset featuring images of African wildlife, including buffalo, elephant, rhino, and zebras.
• Argoverse: A dataset containing 3D tracking and motion forecasting data from urban environments with rich annotations.
• Brain-tumor: A dataset for detecting brain tumors includes MRI or CT scan images with details on tumor presence, location, and characteristics.
• COCO: Common Objects in Context (COCO) is a large-scale object detection, segmentation, and captioning dataset with 80 object categories.
• COCO8: A smaller subset of the first 4 images from COCO train and COCO val, suitable for quick tests.
• COCO8-Grayscale: A grayscale version of COCO8 created by converting RGB to grayscale, useful for single-channel model evaluation.
• COCO8-Multispectral: A 10-channel multispectral version of COCO8 created by interpolating RGB wavelengths, useful for spectral-aware model evaluation.
• COCO12-Formats: A test dataset with 12 images covering 12 supported image formats (AVIF, BMP, DNG, HEIC, JP2, JPEG, JPG, MPO, PNG, TIF, TIFF, WebP) for validating image loading pipelines.
• COCO16: A subset of the first 16 images from COCO train2017 (8 train + 8 val), suitable for quick tests.
• COCO32: A subset of the first 32 images from COCO train2017 (16 train + 16 val), suitable for quick tests.
• COCO64: A subset of the first 64 images from COCO train2017 (32 train + 32 val), suitable for quick tests.
• COCO128: A smaller subset of the first 128 images from COCO train2017, suitable for tests.
• Construction-PPE: A dataset featuring construction site workers with labeled safety gear such as helmets, vests, gloves, boots, and goggles, including missing-equipment annotations like no_helmet, no_goggle for real-world compliance monitoring.
• Global Wheat 2020: A dataset containing images of wheat heads for the Global Wheat Challenge 2020.
• HomeObjects-3K: A dataset of indoor household items including beds, chairs, TVs, and more—ideal for applications in smart home automation, robotics, augmented reality, and room layout analysis.
• KITTI: A dataset featuring real-world driving scenes with stereo, LiDAR, and GPS/IMU data, used here for 2D object detection tasks such as identifying cars, pedestrians, and cyclists in urban, rural, and highway environments.
• LVIS: A large-scale object detection, segmentation, and captioning dataset with 1203 object categories.
• Medical-pills: A dataset featuring images of medical-pills, annotated for applications such as pharmaceutical quality assurance, pill sorting, and regulatory compliance.
• Objects365: A high-quality, large-scale dataset for object detection with 365 object categories and over 600K annotated images.
• OpenImagesV7: A comprehensive dataset by Google with 1.7M train images and 42k validation images.
• Roboflow 100: A diverse object detection benchmark with 100 datasets spanning seven imagery domains for comprehensive model evaluation.
• Signature: A dataset featuring images of various documents with annotated signatures, supporting document verification and fraud detection research.
• SKU-110K: A dataset featuring dense object detection in retail environments with over 11K images and 1.7 million bounding boxes.
• TT100K: Explore the Tsinghua-Tencent 100K (TT100K) traffic sign dataset with 16,817 street-view images across 221 sign categories for robust detection and classification.
• VisDrone: A dataset containing object detection and multi-object tracking data from drone-captured imagery with over 10K images and video sequences.
• VOC: The Pascal Visual Object Classes (VOC) dataset for object detection and segmentation with 20 object classes and over 11K images.
• xView: A dataset for object detection in overhead imagery with 60 object categories and over 1 million annotated objects.

Adding your own dataset#

If you have your own dataset and would like to use it for training detection models with Ultralytics YOLO format, ensure that it follows the format specified above under "Ultralytics YOLO format". Convert your annotations to the required format and specify the paths, number of classes, and class names in the YAML configuration file.

Port or Convert Label Formats#

COCO Dataset Format to YOLO Format#

You can easily convert labels from the popular COCO dataset format to the YOLO format using the following code snippet:

from ultralytics.data.converter import convert_coco

convert_coco(labels_dir="path/to/coco/annotations/")

This conversion tool can be used to convert the COCO dataset or any dataset in the COCO format to the Ultralytics YOLO format. The process transforms the JSON-based COCO annotations into the simpler text-based YOLO format, making it compatible with Ultralytics YOLO models.

Remember to double-check if the dataset you want to use is compatible with your model and follows the necessary format conventions. Properly formatted datasets are crucial for training successful object detection models.

What's Next#

With your dataset formatted, start training your model. Not sure which pretrained model to start from? Compare options in the YOLO26 model family.

FAQ#