Reproduce “GaussianFusion: Unified 3D Gaussian Representation for Multi-Modal Fusion Perception” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiao Zhao, Chang Liu, Mingxu Zhu, Zheyuan Zhang, Linna Song, Qingliang Luo, Chufan Guo, Kuifeng Su
GaussianFusion introduces a unified 3D Gaussian representation for multimodal fusion in autonomous driving, addressing limitations of discrete Bird's-Eye View (BEV) methods. The framework achieves state-of-the-art performance on nuScenes for 3D object detection (74.0 NDS, +2.6 over BEVFusion) and 3D semantic occupancy prediction (28.65 mIoU, +1.11 over OccFusion), demonstrating improved accuracy and efficiency.
Key insights
- A continuous 3D Gaussian representation effectively preserves fine-grained details and adaptively models uncertainty, facilitating superior multimodal alignment and interaction compared to discrete BEV grids.
- Principled Gaussian initialization and an iterative offset-based updating mechanism for Gaussian properties are critical for robust fusion and allow for gradual reduction of inter-modal disparities.
- Incorporating Gaussian shape priors into deformable attention within a shared encoder enhances the learning of uncertain complementary features, leading to more accurate object representations and improved overall perception performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.