Out-of-Distribution Semantic Occupancy Prediction

To address the issue of insufficient detection of long-tail distributions and abnormal objects in current autonomous driving methods, researchers proposed the Out-of-Distribution Semantic Occupancy Prediction task. To this end, two datasets containing synthetic anomalies were constructed: VAA-KITTI and VAA-KITTI-360. The OccOoD framework, which integrates Cross-Space Semantic Refinement (CSSR), was also developed. Experiments showed that this framework achieved AuROC of 65.50% and AuPRCr of 31.83% within a radius of 1.2 million meters, significantly improving the detection sensitivity and generalization ability for unknown obstacles. The relevant datasets and source code have been made available on the GitHub platform.