Kinect 2010: depth cameras and real-time pose estimation at home

Microsoft launched Kinect in North America on 4 November 2010. Real-time pose estimation used random forests on depth images at Xbox 360 speeds, and the accessory sold eight million units in sixty days.

Event date: Published:
HistoryResearch

Microsoft released Kinect for Xbox 360 in North America on 4 November 2010. The accessory used a depth camera stack derived from PrimeSense technology to map living rooms in 3D and let players control games with body motion instead of a handheld pad.

Random forests at console speed

Microsoft Research Cambridge and Xbox Incubation built real-time skeletal tracking by classifying each depth pixel into body parts with random forests of decision trees, following Leo Breiman’s 2001 ensemble recipe (article). Jamie Shotton, Andrew Fitzgibbon, Andrew Blake, and colleagues trained on hundreds of thousands of synthetic depth images (synthetic training data) so the model generalized to real homes. The system ran in under five milliseconds per frame (about 200 frames per second) on the Xbox 360 GPU, enabling stable pose estimation for games and later research tools.

The CVPR 2011 paper won best paper at the conference and cited Breiman’s forests explicitly. Alex Kipman led the broader Kinect product effort at Microsoft.

A consumer electronics record

Kinect sold about eight million units in sixty days, which Guinness World Records listed as the fastest-selling consumer electronics device at the time. The launch brought supervised learning on depth data into millions of households years before phone-based AR went mainstream.

Why it matters

Kinect showed that classical ensembles plus GPU throughput could ship in a $150 sensor bundle, not only in research labs. It linked random forests to embodied machine learning and foreshadowed depth sensing in robotics and telepresence, while deep learning later took over many vision tasks AlexNet helped ignite.