
MR
Mohammad Rashid
· 1 min read
ResearcharXiv cs.CV
Shared Mobile AR Without Depth Sensors from Monocular Point Maps
arXiv:2611.18679v1 Announce Type: new
Abstract: Depth-aware augmented reality on phones needs dense geometry of the surrounding room. iPhones and iPads without LiDAR expose only sparse feature points and detected planes, and ARKit often takes tens of seconds to report the first wall. We describe a scan-free pipeline for these devices. A point-map network (MoGe-2) predicts metric geometry from each camera frame, the remaining depth shift is recovered from the camera intrinsics, and each frame's scale is aligned to views already in the map. A persistent voxel map with visibility-based expiry answers occlusion and collision queries. Co-located devices share one coordinate frame by solving yaw and floor-plane translation from sightings of each other. We evaluate the deployed system on 309 device-sessions from 19 device models without LiDAR, with 179 device-sessions from LiDAR phones in the same deployment as a reference population. A first dense depth map was recorded within 10 s in 80% of sessions, while ARKit had reported a wall plane by then in 25%. The median session scale was 0.98 against ARKit's wall planes and 1.00 against its camera translation. Scale changed by a median of 9% between nearby frames, against 0.4% with LiDAR, and depth points spread three times as widely around walls. In multi-device sessions, 69% of devices without LiDAR reached a converged shared frame, against 84% of LiDAR phones. The main remaining error is scale that changes from frame to frame.
Original source
This story was published by arXiv cs.CV and written by Mohammad Rashid. SyncAI.news shows a preview; the complete article is on the publisher's site.
Read the full story on arxiv.org


