4DR360 proposes state reasoning for joint 3D detection and occupancy prediction in 4D radar-camera perception

The authors propose 4DR360, a framework for 360-degree full-scene perception that jointly performs 3D object detection and semantic occupancy prediction using 4D millimeter-wave radar and camera data. The method models semantic occupancy as a persistent scene state propagated through stages for coarse-to-fine feature aggregation.

• State-guided BEV Enhancement (SBE) strengthens intra-frame Bird's Eye View representation.
• Doppler-guided Temporal Fusion (DTF) preserves state evidence over longer temporal horizons.
• The ManTruckScenes dataset is extended with satellite-map-based generated occupancy labels.
• Evaluation uses a unified cross-dataset protocol pairing ManTruckScenes with OmniHD-Scenes.

The framework addresses the limitation of existing methods that optimize detection while decoding boxes and occupancy with limited interaction, aiming to advance radar-based multi-task learning.