unidepth v2 vitl14

UniDepthV2 is a universal monocular metric depth estimation model that predicts metric depth from a single RGB image, operating without camera intrinsics and generalizing across diverse indoor and outdoor scenes.

Capabilities

The model directly outputs depth in metric units (e.g., meters) from any single image, eliminating the need for per-scene calibration or transformation. It uses a Vision Transformer (ViT-L/14) backbone and is designed for zero-shot cross-dataset performance.

Benchmark Results

At the time of submission, UniDepthV2 achieved first place on the KITTI depth prediction benchmark. It also holds state-of-the-art results on the NYU Depth V2 and KITTI Eigen test sets, as recorded on Papers With Code.

Visual Examples

References

• UniDepthV2 paper: arXiv 2502.20110
• UniDepthV1 paper (CVPR 2024): arXiv 2403.18913
• Project page: lpiccinelli-eth.github.io/pub/unidepth/