From the source
Apple researchers present Depth Pro 2, a single network that estimates dense metric depth for both images and video at full resolution, and release a family of models with code and weights.
- The team says prior monocular depth methods trade off temporal consistency against boundary sharpness, with multi-frame methods favoring consistency and single-frame methods favoring sharpness.
- Depth Pro 2 is trained as one scalable network that processes images and video and predicts depth maps at native resolution, which the authors say lets it be metrically accurate, sharp and temporally consistent at once.
- Apple says the models outperform prior work on all of those dimensions and define a new Pareto frontier in metric accuracy, sharpness, temporal consistency and runtime.
- The release follows Depth Pro, a zero-shot metric monocular depth foundation model that Apple described in October 2024 as producing a 2.25-megapixel depth map in 0.3 seconds on a standard GPU.