Research map / 3D vision / Depth and stereo

Depth Pro: Sharp Monocular Metric Depth in Less Than a Second

ICLR 2025 · 3D vision › Depth and stereo · approach: Self-supervised monocular depth · 20 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Depth Pro is a zero-shot metric monocular depth foundation model that uses a multi-scale plain ViT with a two-stage real/synthetic training curriculum and new matting/segmentation-based boundary metrics, producing 2.25MP sharp depth in 0.3s, plus a focal length head.

Problem

Zero-shot monocular depth estimation that is metric (absolute scale) without camera intrinsics, produces sharp high-resolution boundaries, and runs at low latency.

Approach

Applies a shared plain ViT encoder to multi-scale patches of a fixed 1536x1536 input, fuses them with a global image encoder and DPT-style decoder; trains with a curriculum that moves from mixed real+synthetic data to synthetic-only with derivative losses for sharp boundaries; evaluates boundaries using matting/segmentation masks; and estimates focal length from frozen depth features.

Key results

• Depth Pro achieves the best average rank (2.5) across six zero-shot metric depth datasets, versus 3.7 for the next best (Metric3D v2, gray-marked).
• Depth Pro reaches boundary F1 of 0.409 on Sintel and recall of 0.173 on AM-2k, higher than all absolute and relative baselines.
• Depth Pro produces a 2.25MP depth map in about 0.3 s on a V100 GPU (341 ms at all tested resolutions), and on PPR10K 64.6% of focal length estimates fall within 25% error versus 34.6% for SPEC.

Measured effects reported in the paper

Surprising findings

• Predicting at high native resolution is necessary but not sufficient for accurate boundaries; PatchFusion and ZeroDepth, which predict at full input resolution, trail lower-resolution methods.
• Inverting the usual order, the second training stage uses only synthetic data to sharpen boundaries after mixed real+synthetic training.

Limitations

• Limited on translucent surfaces and volumetric scattering where pixelwise depth is ill-posed.
• Some boundary metrics rely on segmentation masks that may include non-occluding boundaries, requiring manual filtering.
• Zero-shot comparisons mix systems trained on different datasets, and some baselines are not strictly zero-shot.

Experimental setup

Datasets: Keystone HRWSI RedWeb TartanAir Hypersim Sintel Spring iBims AM-2k P3M-10k DIS-5k Booster

Benchmarks: Booster ETH3D Middlebury NuScenes Sintel Sun-RGBD Spring iBims AM-2k P3M-10k DIS-5k DDDP

Metrics: delta_1 delta_2 delta_3 AbsRel Log10 SI-Log point-cloud metrics boundary F1 boundary recall delta_25%

Baselines: Depth Anything v2 Metric3D v2 Marigold PatchFusion UniDepth ZoeDepth ZeroDepth SPEC im2pcl

Builds on: DPT DINOv2 MiDaS Metric3D

Closest papers by meaning

• SharpDepth: Sharpening Metric Depth Predictions Using Diffusion Distillation (CVPR 2025 · similarity 0.90)
• Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.90)
• A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · similarity 0.90)
• Zero-shot Depth Completion via Test-time Alignment with Affine-invariant Depth Prior (AAAI 2025 · similarity 0.90)
• DepthMaster: Unified Monocular Depth Estimation for Perspective and Panoramic Images (NeurIPS 2026 · similarity 0.89)
• Depth Anything V2 (NeurIPS 2024 · similarity 0.89)