A novel multimodal 3D object detection algorithm, IS-Fusion, is presented in this work. It enhances the previous multimodal BEV representation by incorporating instance-level information, which significantly improves detection performance. IS-Fusion achieves the best performance among the published works to date.

We propose a new collaborative 3D object detection algorithm (DI-V2X) aimed at minimizing the domain discrepancy of the input data from various sources, including vehicles and infrastructure. Our approach features three key modules, the Domain-Mixing Instance Augmentation (DMA) module, the Progressive Domain-Invariant Distillation (PDD) module, and the Domain-Adaptive Fusion (DAF) module. DI-V2X achieves the best performance in V2X perception.

We present a novel learning paradigm, LWSIS, that inherits the fruits of off-the-shelf 3D point cloud to guide the training of 2D instance segmentation models to save mask-level annotations. We advocate a new dataset nuInsSeg based on nuScenes to extend existing 3D LiDAR annotations with 2D image segmentation annotations.

We prsent SSDA3D in this work, which is the first effort for semi-supervised domain adaptation in the context of 3D object detection. SSDA3D is achieved by a novel framework that jointly addresses inter-domain adaptation and intra-domain generalization.

We propose a proposal-level point cloud SSL framework, named ProposalContrast, that conducts proposal-wise contrastive pre-training for detection-aligned representation learning. We comprehensively demonstrate the generalizability of our model across diverse 3D detector architectures and datasets.