Google Objectron Dataset

Load Objectron fast with one line of code. 15K annotated video clips and 4M annotated images. Stream Objectron while training models in PyTorch & TensorFlow.

Visualization of the Objectron bike training set dataset in the Activeloop Platform
Objectron dataset
What is Objectron Dataset?
The Objectron dataset consists of several short, object-centric video clips where the camera steadily moves around the object and captures it from different angles. The videos contain manually annotated 3D bounding boxes describing the object's position, orientation, and dimensions. The dataset also comes with the metadata from AR sessions, including camera poses, sparse point clouds, and characterization of the planar surfaces from the surrounding environment. The dataset is collected from 10 countries across five continents, thus ensuring geo-diversity. It consists of 17,095 object instances with 14,819 annotated video clips complemented with 4M annotated images in the following categories: bikes, books, cameras, cereal boxes, chairs, cups, laptops, and shoes.
Download Objectron Dataset in Python
Instead of downloading the Objectron dataset in Python, you can effortlessly load it in Python via our open-source package Hub with just one line of code.
Load Objectron Dataset Training Subset for the Category Book in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_book_train")
Copied!
Load Objectron Dataset Testing Subset for the Category Book in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_book_test")
Copied!
Load Objectron Dataset Training Subset for the Category Bike in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_bike_train")
Copied!
Load Objectron Dataset Testing Subset for the Category Bike in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_bike_test")
Copied!
Load Objectron Dataset Training Subset for the Category Bottle in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_bottle_train")
Copied!
Load Objectron Dataset Testing Subset for the Category Bottle in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_bottle_test")
Copied!
Load Objectron Dataset Training Subset for the Category Camera in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_camera_train")
Copied!
Load Objectron Dataset Testing Subset for the Category Camera in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_camera_test")
Copied!
Load Objectron Dataset Training Subset for the Category Cereal box in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_cereal_box_train")
Copied!
Load Objectron Dataset Testing Subset for the Category Cereal box in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_cereal_box_test")
Copied!
Load Objectron Dataset Training Subset for the Category Chair in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_chair_train")
Copied!
Load Objectron Dataset Testing Subset for the Category Chair in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_chair_test")
Copied!
Load Objectron Dataset Training Subset for the Category Cup in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_cup_train")
Copied!
Load Objectron Dataset Testing Subset for the Category Cup in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_cup_test")
Copied!
Load Objectron Dataset Training Subset for the Category Shoe in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_shoe_train")
Copied!
Load Objectron Dataset Testing Subset for the Category Shoe in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_shoe_test")
Copied!
Load Objectron Dataset Training Subset for the Category Laptop in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_laptop_train")
Copied!
Load Objectron Dataset Testing Subset for the Category Laptop in Python
1
import hub
2
ds = hub.load("hub://activeloop/objectron_laptop_test")
Copied!
Objectron Dataset Structure
Objectron Data Fields
camera_extrinsics: tensor containing row major 4x4 transformation matrix describing the camera pose w.r.t. the world origin. The world origin is where the AR session has started.
camera_intrinsics: tensor containing row-major 3x3 intrinsic matrix describing the focal length and the principal point of the camera.
camera_projection: tensor containing row major 4x4 projection matrix.
camera_view: tensor containing row major 4x4 view matrix.
object_orientation: tensor containing string indicating the orientation of the camera (portrait, landscape).
object_scale: tensor containing list of object scales.
object_translation: tensor containing list of object translations.
object_visibility: tensor containing list of annotation visibilities.
plane_center: tensor containing center point for the ground plane where objects are sitting on.
plane_normal: tensor containing the normal vector to the ground plane where objects are sitting on.
point_2d: tensor containing float numbers of all 2D points.
point_3d: tensor containing float numbers of all 3D points.
point_num: tensor containing point numbers for each instance. The bounding box has 9 points, skeleton may have a varied number of points.
image_channels: an integer representing number of channels in the image.
image_height: an integer representing height of the image.
image_id: an integer representing frame number in the sequence.
image_timestamp: an integer representing the microsecond timestamp of the video frame in the video stream.
image_width: an integer representing height of the image.
instance_num: an integer representing number of object instances in this frame.
image: tensor containing an image of dimension (image_height x image_width x image_channels).
Objectron Data Splits
​Google Objectron dataset training set for class book is composed of 460838 images. 
​Google Objectron dataset testing set for class book is composed of 115389 images.
​Google Objectron dataset training set for class bike is composed of 123675 images. 
​Google Objectron dataset testing set for class bike is composed of 30714 images.
​Google Objectron dataset training set for class bottle is composed of 378932 images. 
​Google Objectron dataset testing set for class bottle is composed of 93888 images.
​Google Objectron dataset training set for class camera is composed of 179175 images. 
​Google Objectron dataset testing set for class camera is composed of 45939 images.
​Google Objectron dataset training set for class cereal_box is composed of 314910 images.
​Google Objectron dataset testing set for class cereal_box is composed of 76827 images.
​Google Objectron dataset training set for class chair is composed of 394529 images.
​Google Objectron dataset testing set for class chair is composed of 97621 images.
​Google Objectron dataset training set for class cup is composed of 436513 images.
​Google Objectron dataset testing set for class cup is composed of 109870 images.
​Google Objectron dataset training set for class shoe is composed of 445919 images.
​Google Objectron dataset testing set for class shoe is composed of 115019 images.
​Google Objectron dataset training set for class laptop is composed of 388792 images.
​Google Objectron dataset testing set for class laptop is composed of 97074 images.
How to use Objectron Dataset with PyTorch and TensorFlow in Python
Train a model on Objectron Dataset with PyTorch in Python
Let's use Hub's built-in PyTorch one-line dataloader to connect the data to the compute:
1
dataloader = ds.pytorch(num_workers=0, batch_size=4, shuffle=False)
Copied!
Train a model on Objectron dataset with TensorFlow in Python
1
dataloader = ds.tensorflow()
Copied!
Additional Information about Objectron Dataset
Objectron Dataset Description
Homepage: https://www.objectron.dev/​
Repository: https://github.com/google-research-datasets/Objectron​
Paper: https://arxiv.org/abs/2012.09988​
Point of Contact: [email protected], [email protected]
Objectron Dataset Curators
Adel Ahmadyan, Liangkai Zhang, Jianing Wei, Artsiom Ablavatski, Matthias Grundmann
Objectron Dataset Licensing Information
​Computational Use of Data Agreement 1.0 (C-UDA-1.0) license
Objectron Dataset Citation Information
1
@article{objectron2021,
2
  title={Objectron: A Large Scale Dataset of Object-Centric Videos in the Wild with Pose Annotations},
3
  author={Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, Matthias Grundmann},
4
  journal={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
5
  year={2021}
6
}
Copied!
Objectron Dataset FAQs
What is the Objectron Dataset for Python?
The Objectron Dataset was introduced to advance state of the art in 3D object detection. It consists of object-centric 15K annotated videos and 4M annotated images.
What is the Objectron Dataset used for?
The Objectron Dataset is commonly used to improve 3D shape representation and enable new research and applications in 3D understanding, video models, object retrieval, view synthetics, and 3D reconstruction.
What are the Google Objectron Dataset classes?
The Objectron Dataset consists of 9 different classes: bikes, books, cameras, cereal boxes, chairs, cups, laptops, and shoes.
 How to download the Objectron Dataset in Python?
You can load Objectron Dataset fast with one line of code using the open-source package Activeloop Hub in Python. See detailed instructions on how to load Objectron Dataset training subset or Objectron Dataset testing subset in Python.
How can I use Objectron Dataset in PyTorch or TensorFlow?
You can stream Objectron Dataset while training a model in PyTorch or TensorFlow with one line of code using the open-source package Activeloop Hub in Python. See detailed instructions on how to ​train a model on Objectron Dataset with Pytorch in Python or train a model on Objectron Dataset with TensorFlow in Python.
Hub community member Manas Gupta has contributed to this dataset documentation. You're awesome, Manas!