AsyncBEV: Cross-modal Flow Alignment in Asynchronous 3D Object Detection
AsyncBEV is a trainable module designed to make 3D Bird's Eye View (BEV) object detection models robust against sensor timing mismatches. It addresses issues where sensors operate at different frequencies, causing errors in dynamic object detection. By estimating and warping the feature flow between asynchronous modalities, AsyncBEV significantly improves performance, especially for moving objects, outperforming existing baselines by substantial margins under various time offsets.