Depth Pro 2

Sharp and Temporally Consistent Metric Monocular Depth

Peiyun Hu, Marko Mihajlovic, Lyndon Duong, Lahav Lipson, Wei Dong,
Amaël Delaunoy, Stephan R. Richter, Vladlen Koltun

Apple

Paper Code Models
Source Depth Pro 2

Depth Pro 2 estimates sharp, metric, temporally consistent depth for images and video at full resolution.

Video comparisons

Select a video to compare

Image comparisons

Sharpness versus temporal consistency

Depth Pro 2 outperforms prior work operating on a single frame or multiple frames on sharpness (Boundary F1 score) and temporal consistency (Jitter). Results averaged over Sintel and Spring datasets.

Method

Architecture: a multi-view transformer processes the input frames jointly, and a DPT decoder predicts an inverse depth map per frame. Input frame 1 Input frame 1 Output frame 1 Output frame 1 Input frame 2 Input frame 2 Output frame 2 Output frame 2 Input frame N Input frame N Output frame N Output frame N ⋮ ⋮ Multi-view transformer 1 2 DPT Decoder DPT Decoder DPT Decoder Tokens frame i−1 Tokens frame i Tokens frame i+1 1 Attention within frame Attention across frames Dense feature map Layer L1 H/32×W/32 Layer L2 H/16×W/16 Layer L3 H/8×W/8 Layer L4 H/4×W/4 2 DPT Decoder

Our depth estimation network consists of a multi-view transformer that ingests one or multiple input frames and processes them jointly, and a DPT decoder, which predicts an inverse depth map per input frame. 1 The transformer shares information across frames: after the first third of its blocks, it alternates attention within each frame and attention across frames. 2 The DPT decoder takes tokens from several transformer blocks, assembles them into feature maps at 1/32 to 1/4 of the input resolution, and decodes them into an inverse depth map per frame.

Citation

@article{DepthPro2:2026:arxiv,
  title      = {Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth},
  author     = {Peiyun Hu and Marko Mihajlovic and Lyndon Duong and Lahav Lipson and Wei Dong and Ama\"{e}l Delaunoy and Stephan R. Richter and Vladlen Koltun},
  journal    = {arXiv:261X.XXXXX},
  year       = {2026},
  url        = {https://arxiv.org/abs/261X.XXXXX},
}