Four million frames of people circling their shoes

What it does

Objectron is a dataset of roughly 15,000 short video clips and 4 million images showing everyday objects—bikes, bottles, chairs, shoes, and others—from multiple angles as the camera circles them. Each clip carries manually annotated 3D bounding boxes describing the object’s position, orientation, and dimensions, plus AR session metadata like camera poses, sparse point clouds, and surrounding planes. The raw data weighs in at 1.9 TB (4.4 TB total with processed records), and the repository provides schemas, parsing scripts, and notebooks to load it into Tensorflow, Jax, or Pytorch.

The interesting bit

The footage was collected across ten countries on five continents, so your model gets to learn on furniture and footwear from wildly different living rooms. Google also trained MediaPipe’s 3D object detection models for shoes, chairs, mugs, and cameras on this data, which gives you a concrete reference for annotation quality.

Key highlights

• 15K annotated videos and 4M images across nine household categories
• Every sample includes high-res imagery, camera pose, object pose, point clouds, and planar surfaces
• Pre-shuffled tf.record shards and tf.SequenceExample files for Tensorflow and Pytorch pipelines
• Evaluation scripts using 3D IoU for oriented bounding boxes
• Apache Beam jobs and notebooks for parsing raw protobuf annotations and even training NeRF models

Caveats

• The repository is explicitly not an officially supported Google product.
• The dataset lives in a public Google Cloud bucket; plan for 1.9 TB of raw downloads plus processed records.
• The companion MediaPipe models cover only four of the nine dataset categories: shoes, chairs, mugs, and cameras.

Verdict

Grab this if you’re building 3D detectors, pose estimators, or NeRFs and need real-world multi-view data with ground-truth geometry. Skip it if you were hoping for a lightweight drop-in model—this is the raw material, not the finished product.