Apple Releases Depth Pro 2: Single Network Achieves Sharp, Consistent Metric Depth for Images and Video
Apple releases Depth Pro 2, a single network that resolves the trade-off between temporal consistency and boundary sharpness in monocular depth estimation.
Apple's machine learning research team has released Depth Pro 2, a new approach to estimating dense metric depth from images and video. The model is designed to overcome a persistent limitation in prior monocular depth estimation methods: the trade-off between temporal consistency and boundary sharpness. Multi-frame methods typically favor consistency across frames but lose edge definition, while single-frame methods produce sharp boundaries but struggle with stability over time.
Depth Pro 2 addresses this by training a single, scalable network capable of processing both static images and video sequences at their native full resolution. According to Apple, the resulting depth maps are simultaneously metrically accurate, sharp at object edges, and temporally consistent from frame to frame. The company states that the model outperforms prior work in all these dimensions, defining a new Pareto frontier in terms of accuracy, sharpness, consistency, and runtime.
The release includes code and model weights, making the technology available for external evaluation and integration. This follows Apple's previous Depth Pro model, which focused on fast, high-resolution metric depth for single images. Depth Pro 2 extends this capability to video without sacrificing the sharpness or speed characteristics of its predecessor.