M squared-Occ: Resilient 3D Semantic Occupancy Prediction for Autonomous Driving with Incomplete Camera Inputs
M2-Occ addresses unreliable camera inputs by developing a framework to predict 3D occupancy and semantics even when views are missing. It uses Multi-view Masked Reconstruction to recover geometry from neighboring cameras and a Feature Memory Module with learnable prototypes to regularize semantic predictions. This improves robustness in real-world driving scenarios where complete sensor coverage is often impossible.