Ultralytics YOLO26: depth estimation and export tutorial

Ultralytics introduces YOLO26 depth models that predict a distance in meters for each pixel from a single RGB image. The team describes five sizes, from nano to extra large, and demonstrates indoor and outdoor scenes without a stereo camera or LiDAR input.

The walkthrough uses Ultralytics Platform to upload the eight-image Depth 8 dataset, inspect depth overlays, and train a nano model for 100 epochs. The presenter cautions that this small demonstration needs more data and longer training for production. Export examples cover Core ML, OpenVINO, ONNX and TensorRT for deployment on users' own hardware.

An iPhone demonstration shows on-device inference in the public Ultralytics YOLO app, which the speakers also say is available for Android. They report about 20 milliseconds of inference for a YOLO26M model at 640 resolution on the demonstrated iPhone; this is a session result, not a guarantee for other devices.

For a new camera or environment, the team recommends testing pretrained predictions first, then calibrating with roughly 50 to 100 images paired with sensor measurements before considering further training. They acknowledge limited fisheye testing and difficulties with glass and water. Stereo inputs and 3D object detection are discussed as future work, while precision measurement remains a reason to retain dedicated sensors.