A new detection system for self-driving cars stops assuming camera and LiDAR data are equally trustworthy at every moment - and lets a corrupted feed get outvoted instead of dragging down the whole prediction.
Researchers built SARFusion, a 3D object detection model that splits its decision-making into three parallel paths: one reading only camera data, one reading only LiDAR, and one fusing both. A "Scene Reliability Prior," estimated from the broader driving scene, combines with object-level evidence to route each individual detection query to whichever path is most trustworthy at that moment. On the nuScenes benchmark, the system scored 72.5 mean average precision and 74.4 on the NDS accuracy metric, and held up under simulated sensor corruption and changing environmental conditions.
The pitch is a fix for a real weak spot in current fusion detectors: when camera and LiDAR data get blended into one tightly coupled representation, a degraded sensor - fog on a lens, a dropout in the point cloud - can poison the fused output rather than just weakening one input. Routing decisions per-query and per-scene, instead of per-model, is a more surgical way to isolate that damage.
Still, this is a benchmark result on nuScenes, not a test on an actual foggy highway - and self-driving papers have a long history of clean numbers that get messier once they leave the dataset.