A new geometry back-end called DH-Active extends iPhone LiDAR's useful range without training a single model.
The LiDAR scanner on recent iPhones is fast but shallow — returns get sparse and short-range quickly. DH-Active, described in a new arXiv paper, sidesteps that ceiling by treating the sensor as a scale reference rather than a depth map. Near-field LiDAR returns anchor the metric pose between two camera views; points the sensor misses are then recovered through triangulation. Where the geometry is too shaky to trust, the system explicitly abstains and flags the gap rather than guessing. The core pipeline runs at 1.11 ms median latency on CPU, roughly 38 times faster than a DINOv2-L visual model running on GPU.
The practical stakes are real. On-device depth estimation is a bottleneck for AR, accessibility tools, and spatial computing — all areas Apple has staked product bets on. A training-free approach at near-millisecond CPU cost is a meaningful constraint for shipping on hardware with strict thermal and power budgets. By explicitly marking uncertain regions instead of papering over them, DH-Active also gives downstream applications something to act on: a confidence signal, not a hallucinated surface.
The paper is careful about what it is not claiming. A 1.26-billion-parameter learned model still wins on accuracy once scale is aligned externally — so this is not a replacement for neural depth estimation, it is a lightweight complement. Several alternatives the authors tried, including defocus fusion and ICP over visual-inertial tracks, failed outright, which is a useful accounting of the problem's difficulty.