Objectron

Why we included this project

This is the dataset behind Google's MediaPipe object detection models for shoes, chairs, mugs, and cameras, and it still holds up as a reference when you are training or benchmarking models that predict object pose in the real world. It packs about 15,000 short object-centric video clips and over four million annotated images, with each sample carrying camera pose, sparse point-cloud, plane, and 3D bounding box data. Because the footage was gathered across ten countries on five continents, it offers geographic variety that many single-lab datasets lack. The repo includes ready-to-use tf.record formats and a notebook for downloading the data, so you can feed it straight into a TensorFlow or PyTorch training loop without much setup. Anyone working on 3D object detection, pose estimation, or AR-centric perception will find the annotations and evaluation conventions here worth studying.