Objectron Dataset

Objectron is a dataset of short object centeric video clips with pose annotations.

objectron_samples

The Objectron dataset is a collection of short, object-centric video clips, which are accompanied by AR session metadata that includes camera poses, sparse point-clouds and characterization of the planar surfaces in the surrounding environment. In each video, the camera moves around the object, capturing it from different angles. The data also contain manually annotated 3D bounding boxes for each object, which describe the object’s position, orientation, and dimensions. The dataset consists of 15K annotated video clips supplemented with over 4M annotated images in the following categories: bikes, books, bottles, cameras, cereal boxes, chairs, cups, laptops, and shoes. In addition, to ensure geo-diversity, our dataset is collected from 10 countries across five continents. Along with the dataset, we are also sharing a 3D object detection solution for four categories of objects — shoes, chairs, mugs, and cameras. These models are trained using this dataset, and are released in MediaPipe, Google's open source framework for cross-platform customizable ML solutions for live and streaming media.

Key Features

  • 15000 annotated videos and 4M annotated images
  • All samples include high-res images, object pose, camera pose, point-cloud, and surface planes.
  • Ready to use examples in various tf.record formats, which can be used in Tensorflow/PyTorch.
  • Object-centric multi-views, observing the same object from different angles.
  • Accurate evaluation metrics, like 3D IoU for oriented 3D bounding boxes.

Dataset Format

The data is stored in the objectron bucket on Google Cloud storage. Check out the Download Data notebook for a quick review of how to download/access the dataset. The following assets are available:

  • The video sequences (located in /videos/class/batch-i/j/video.MOV files)
  • The annotation labels containing the 3D bounding boxes for objects. The annotation protobufs are located in /videos/class/batch-i/j/geometry.pbdata files. They are formatted using the object.proto. See [example] on how to parse the annotation files.
  • AR metadata (such as camera poses, point clouds, and planar surfaces). They are based on a_r_capture_metadata.proto. See example on how to parse these files.
  • Processed dataset: sharded and shuffled tf.records of the annotated frames, in tf.example format and videos in tf.SequenceExample format. These are used for creating the input data pipeline to your models. These files are located in /v1/records_shuffled/class/ and /v1/sequences/class/.
  • Supporting scripts to run evaluation based on the 3D IoU metric.
  • Supporting scripts to load the data into Tensorflow, Jax and Pytorch and visualize the dataset, including “Hello World” examples.
  • Supporting Apache Beam jobs to process the datasets on Google Cloud infrastructure.
  • The index of all available samples, as well as train/test splits for easy access and download.

Raw dataset size is 1.9TB (including videos and their annotations). Total dataset size is 4.4TB (including videos, records, sequences, etc.). This repository provides the required schemas and tools to parse the dataset.

class bike book bottle camera cereal_box chair cup laptop shoe
#videos 476 2024 1928 815 1609 1943 2204 1473 2116
#frames 150k 576k 476k 233k 396k 488k 546k 485k 557k

GitHub