Co-Fusion4D: Spatio-temporal Collaborative Fusion for Robust 3D Object Detection

Co-Fusion4D is a new framework for 3D object detection in autonomous driving that addresses spatiotemporal inconsistencies in Bird's Eye View (BEV) detectors by using current-frame-centric fusion with historical frame alignment. The approach achieves state-of-the-art performance on the nuScenes benchmark (74.9% mAP, 75.6% NDS) through a Dual Attention Fusion module that enhances temporal stability without test-time augmentation.

Co-Fusion4D represents a meaningful advancement in autonomous driving perception systems, addressing a fundamental technical challenge that has constrained BEV-based 3D object detectors. The core problem—temporal feature misalignment caused by object and ego-motion across video frames—directly impacts safety-critical decision-making in autonomous vehicles. Traditional approaches either accumulate errors through uniform multi-frame fusion or fail to maintain consistency across time, creating unreliable perception during dynamic driving scenarios.

The framework's current-frame-centric strategy with selective historical incorporation diverges from conventional multi-frame fusion paradigms. By treating the current frame as primary information and filtering historical frames through spatiotemporal alignment before fusion, Co-Fusion4D reduces cumulative drift while extracting valuable temporal cues. This dominant-complementary mechanism addresses a practical engineering challenge: balancing temporal richness against alignment accuracy.

The Dual Attention Fusion module further refines feature interaction by simultaneously processing intra-frame spatial relationships and inter-frame temporal relationships. This adaptive approach enables the detector to emphasize motion-consistent regions while suppressing spurious correlations that emerge from misaligned features. The benchmark results demonstrate tangible performance gains without external training data or computationally expensive test-time augmentation, suggesting practical deployment feasibility.

For autonomous driving development, improved 3D object detection directly translates to safer perception pipelines. The absence of dependency on external data or augmentation strategies makes this approach particularly relevant for commercial deployment where computational efficiency and data governance matter significantly. As autonomous driving companies optimize their perception stacks, such algorithmic improvements in fundamental detection reliability accumulate into meaningful safety enhancements across the industry.

• →Co-Fusion4D solves temporal feature misalignment in BEV-based 3D object detectors through current-frame-centric fusion with spatiotemporal filtering.
• →Dual Attention Fusion module adaptively combines spatial and temporal attention to emphasize motion-consistent regions while suppressing noise.
• →Achieves 74.9% mAP and 75.6% NDS on nuScenes without test-time augmentation or external training data, indicating practical deployment feasibility.
• →Current-frame-dominant approach mitigates cumulative alignment errors that plague conventional uniform multi-frame fusion strategies.
• →Framework targets a core safety-critical problem in autonomous driving where temporal inconsistency directly impacts perception reliability.