Tao Train Depth Anything V2

NVIDIA's TAO monocular depth estimation skill wraps both Metric and Relative Depth Anything v2 architectures to predict per-pixel depth from single RGB images. The workflow requires careful pairing of model type and dataset class based on your data encoding (disparity pixels versus metric meters), and you'll spend time getting annotation files right since the convert utility is picky about directory patterns and extension swaps. AutoML is on by default for training, which is helpful but adds routing complexity. The skill handles the full pipeline from training through TensorRT deployment, though deploy actions bounce to a separate container. If you're working with NYU depth data in raw uint16 millimeters, use the NYUDV2 dataset classes to avoid silent NaN losses from unit conversion failures.

Install to Claude Code

npx -y skills add nvidia/skills --skill tao-train-depth-anything-v2 --agent claude-code