Supplementary for Multi-user, Scalable 3D Object Detection in AR Cloud

Siddharth Choudhary, Nitesh Sekhar, Siddharth Mahendran, Prateek SinghalMagic Leap, CA, USA

{schoudhary,nsekhar,smahendran,psinghal}@magicleap.com

1. Scene SnapshotsFigure 1 shows the ScanNet scene snapshots when us-

ing centralized approach with 1 user or distributed approachwith 10 or 50 users. First column shows results of the cen-tralized approach with 1 user. 2nd column shows resultsof the distributed approach with 10 users and the 3rd col-umn with 50 users. Each row represents a different Scan-Net scene. Ground-truth boxes are shown in blue whereasestimated boxes are shown in red. First three rows containgood predictions whereas the last row includes some erro-neous predictions (sofa in column 2 with 10 users).

1

1 user 10 users 50 users

Figure 1: Comparison of 3D bounding boxes found in a scene while using centralized (1 user) vs distributed (10 and 50 users)approach. First column shows results of the centralized approach with 1 user. 2nd column shows results of the distributedapproach with 10 users and the 3rd column with 50 users. Each row represents a different ScanNet scene. Ground-truthboxes are shown in blue whereas estimated boxes are shown in red. First three rows contain good predictions whereas thelast row includes some erroneous predictions (sofa in column 2 with 10 users).

2