Objectron
Objectron is a large-scale dataset comprising 15,000 annotated short video clips and over 4 million annotated images focused on nine common object categories: bikes, books, bottles, cameras, cereal boxes, chairs, cups, laptops, and shoes. Collected from ten countries across five continents to ensure geographic diversity, each video features a camera moving around and above an object to capture multiple views. The dataset is enriched with augmented reality session metadata, including camera poses, sparse point clouds, and planar surface information from the surrounding environment. Every object within the videos is annotated with a precise 3D bounding box describing its position, orientation, and dimensions. Designed for computer vision research, particularly 3D object detection and pose estimation, Objectron supports machine learning workflows in TensorFlow, PyTorch, and Jax. It includes processed data in tf.record formats for immediate model training, along with tools for visualization and evaluation using m