Depth Anything 3 Small

Depth Anything 3 DA3-Small is a multi-view depth estimation and camera pose estimation model that predicts spatially consistent geometry from arbitrary visual inputs, with or without known camera poses. It uses a single plain transformer backbone with a unified depth-ray representation, trained exclusively on public academic datasets.

Key Capabilities

• Relative depth estimation from single or multiple images
• Camera pose estimation (extrinsics and intrinsics)
• Pose conditioning for geometry-aware inference
• Feed-forward 3D Gaussian estimation for novel view synthesis
• Multi-camera spatial perception (e.g., autonomous driving setups with non-overlapping views)

Performance

DA3-Small significantly outperforms Depth Anything 2 for monocular depth estimation and VGGT for multi-view depth and pose estimation. On the project page, DA3 surpasses VGGT by 35.7% in camera pose accuracy and 23.6% in geometric accuracy (arXiv reports 44.3% and 25.1% respectively). The model also reduces drift in large-scale SLAM applications, outperforming COLMAP (which requires over 48 hours).

Architecture and Training

DA3-Small has 0.08B parameters and uses a teacher-student training paradigm to achieve detail and generalization on par with DA2. A DPT head can be trained on the frozen backbone to predict 3D Gaussian parameters for generalizable novel view synthesis. The model supports DA3-Streaming for ultra-long video sequences using under 12GB GPU memory via sliding-window inference.

Benchmarks

The paper introduces a visual geometry benchmark covering camera pose estimation, any-view geometry, and visual rendering. A benchmark evaluation pipeline is released for pose estimation and 3D reconstruction on five datasets. The model was accepted as an ICLR 2026 Oral.

Limitations

• Trained on academic datasets; performance may vary on domain-specific images
• Results depend on image quality, lighting conditions, and scene complexity