A Wireless Camera Scheduling Framework for Low-Latency 3D Reconstruction

Existing real-time high-quality three-dimensional (3D) reconstruction methods typically assume dense multi-view inputs, which means significant uplink traffic overhead. However, with limited uplink bandwidth in real-world systems, it is difficult to upload all camera views in each uplink phase. This results in sparse and incomplete view coverage during training, ultimately degrading the quality of 3D reconstruction. To tackle this challenge, we propose a resource-aware camera scheduling framework to optimize low-latency 3D reconstruction quality with limited bandwidth. Specifically, we first design a multi-view reconstruction quality prediction network in a multiuser multiple-input multiple-output (MU-MIMO) system, which predicts the reconstruction quality for a set of selected cameras. Experimental results on typical scenes show that the proposed network achieves a mean squared error of 0.003783 in prediction quality. Thus, we can predict the reconstruction quality without executing the complicated reconstruction algorithm. We further develop a camera scheduling algorithm to jointly optimize reconstruction quality and communication latency. Our algorithm reduces the learned perceptual image patch similarity (LPIPS) by 35.66% - 49.98% compared to baseline methods, and can meet the 33.3 ms per-frame latency constraint.