Selected work

Representative Papers

FOMO-3D: Using Vision Foundation Models for Long-Tailed 3D Object Detection

Mar 09, 2026

This work addresses the challenge of detecting rare yet safety-critical 3D objects—such as construction workers—under long-tailed distributions in autonomous driving. The authors propose a multimodal two-stage detection framework that, for the first time, leverages vision foundation models (OWLv2 and Metric3Dv2) to provide semantic and depth priors. A novel camera branch is designed to incorporate these priors, and an attention-based mechanism is employed to fuse LiDAR point cloud features with image features, thereby enhancing both proposal generation and refinement. Experiments on real-world driving data demonstrate that the proposed method significantly improves 3D detection performance on long-tailed categories, validating the effectiveness of integrating vision foundation model priors with multimodal fusion strategies.

1 citationsRead paper

Flux4D: Flow-based Unsupervised 4D Reconstruction

Dec 02, 2025

Existing NeRF and 3D Gaussian Splatting (3DGS) methods for large-scale dynamic 4D reconstruction suffer from limited scalability and reliance on explicit motion annotations, while self-supervised approaches lack cross-scene generalizability and require laborious hyperparameter tuning. Method: We propose the first fully unsupervised, cross-scene trainable 4D Gaussian reconstruction framework—requiring no pretraining, geometric priors, or motion supervision, and taking only multi-view video frames as input. It jointly optimizes the spatial distribution and rigid/non-rigid motion trajectories of 3D Gaussians via photometric consistency and a “prefer-static” regularization, enabling automatic disentanglement of dynamic elements. Results: Our method achieves state-of-the-art performance on outdoor driving datasets, with single-frame inference in seconds. It delivers high-fidelity 4D reconstructions, strong scalability, and superior cross-scene generalization—without scene-specific adaptation.

1 citationsRead paper

Diffusion-guided Generalizable Enhancer for Urban Scene Reconstruction

May 21, 2026

Existing methods for urban scene reconstruction suffer from poor rendering quality under large viewpoint shifts and limited generalization, hindering their applicability in closed-loop simulation. This work proposes GenRe—a diffusion-based universal enhancer that, for the first time, learns cross-scene diffusion priors and efficiently distills them into any pre-trained 3D Gaussian representation without per-scene optimization. GenRe significantly improves rendering fidelity and viewpoint generalization under challenging conditions—such as lane changes—within minutes, outperforming current approaches in both reconstruction quality and computational efficiency. The method effectively supports downstream applications including autonomous driving sensor simulation.

0 citationsRead paper

GenAssets: Generating in-the-wild 3D Assets in Latent Space

Apr 24, 2026

Existing methods struggle to efficiently generate high-quality, complete 3D assets suitable for multi-view simulation from sparse and heavily occluded real-world driving scenes. This work proposes a “reconstruct-then-generate” 3D latent diffusion model that first constructs a high-fidelity object latent space by integrating LiDAR and camera data through occlusion-aware neural rendering trained jointly across multiple scenes. A 3D latent diffusion model is then trained in this space to synthesize diverse assets with complete geometry and appearance. By uniquely combining occlusion-aware neural reconstruction with 3D latent diffusion generation, the method significantly outperforms existing approaches on real-world driving data, achieving superior completeness, diversity, and simulation readiness, thereby enabling large-scale autonomous driving simulation applications.

0 citationsRead paper

Efficient Equivariant Transformer for Self-Driving Agent Modeling

Apr 01, 2026

This work addresses the need for SE(2)-equivariant modeling of multi-agent behaviors in autonomous driving by introducing DriveGATr, a novel architecture that integrates projective geometric algebra with the Transformer framework. The method encodes scene elements as multivectors in the 2D projective geometric algebra ℝ*_{2,0,1} and leverages standard self-attention to implicitly capture geometric relationships without requiring explicit pairwise relative positional encodings, thereby avoiding the quadratic computational overhead typical of conventional approaches. As the first effort to incorporate 2D projective geometric algebra into an equivariant Transformer, DriveGATr achieves traffic simulation performance on par with state-of-the-art methods on the Waymo Open Motion Dataset while offering a superior Pareto trade-off between accuracy and computational efficiency.

0 citationsRead paper

Recent publications

Latest Papers

Diffusion-guided Generalizable Enhancer for Urban Scene Reconstruction

May 21, 2026

Existing methods for urban scene reconstruction suffer from poor rendering quality under large viewpoint shifts and limited generalization, hindering their applicability in closed-loop simulation. This work proposes GenRe—a diffusion-based universal enhancer that, for the first time, learns cross-scene diffusion priors and efficiently distills them into any pre-trained 3D Gaussian representation without per-scene optimization. GenRe significantly improves rendering fidelity and viewpoint generalization under challenging conditions—such as lane changes—within minutes, outperforming current approaches in both reconstruction quality and computational efficiency. The method effectively supports downstream applications including autonomous driving sensor simulation.

0 citationsRead paper

GenAssets: Generating in-the-wild 3D Assets in Latent Space

Apr 24, 2026

Existing methods struggle to efficiently generate high-quality, complete 3D assets suitable for multi-view simulation from sparse and heavily occluded real-world driving scenes. This work proposes a “reconstruct-then-generate” 3D latent diffusion model that first constructs a high-fidelity object latent space by integrating LiDAR and camera data through occlusion-aware neural rendering trained jointly across multiple scenes. A 3D latent diffusion model is then trained in this space to synthesize diverse assets with complete geometry and appearance. By uniquely combining occlusion-aware neural reconstruction with 3D latent diffusion generation, the method significantly outperforms existing approaches on real-world driving data, achieving superior completeness, diversity, and simulation readiness, thereby enabling large-scale autonomous driving simulation applications.

0 citationsRead paper

Efficient Equivariant Transformer for Self-Driving Agent Modeling

Apr 01, 2026

This work addresses the need for SE(2)-equivariant modeling of multi-agent behaviors in autonomous driving by introducing DriveGATr, a novel architecture that integrates projective geometric algebra with the Transformer framework. The method encodes scene elements as multivectors in the 2D projective geometric algebra ℝ*_{2,0,1} and leverages standard self-attention to implicitly capture geometric relationships without requiring explicit pairwise relative positional encodings, thereby avoiding the quadratic computational overhead typical of conventional approaches. As the first effort to incorporate 2D projective geometric algebra into an equivariant Transformer, DriveGATr achieves traffic simulation performance on par with state-of-the-art methods on the Waymo Open Motion Dataset while offering a superior Pareto trade-off between accuracy and computational efficiency.

0 citationsRead paper

FOMO-3D: Using Vision Foundation Models for Long-Tailed 3D Object Detection

Mar 09, 2026

This work addresses the challenge of detecting rare yet safety-critical 3D objects—such as construction workers—under long-tailed distributions in autonomous driving. The authors propose a multimodal two-stage detection framework that, for the first time, leverages vision foundation models (OWLv2 and Metric3Dv2) to provide semantic and depth priors. A novel camera branch is designed to incorporate these priors, and an attention-based mechanism is employed to fuse LiDAR point cloud features with image features, thereby enhancing both proposal generation and refinement. Experiments on real-world driving data demonstrate that the proposed method significantly improves 3D detection performance on long-tailed categories, validating the effectiveness of integrating vision foundation model priors with multimodal fusion strategies.

1 citationsRead paper

Flux4D: Flow-based Unsupervised 4D Reconstruction

Dec 02, 2025

Existing NeRF and 3D Gaussian Splatting (3DGS) methods for large-scale dynamic 4D reconstruction suffer from limited scalability and reliance on explicit motion annotations, while self-supervised approaches lack cross-scene generalizability and require laborious hyperparameter tuning. Method: We propose the first fully unsupervised, cross-scene trainable 4D Gaussian reconstruction framework—requiring no pretraining, geometric priors, or motion supervision, and taking only multi-view video frames as input. It jointly optimizes the spatial distribution and rigid/non-rigid motion trajectories of 3D Gaussians via photometric consistency and a “prefer-static” regularization, enabling automatic disentanglement of dynamic elements. Results: Our method achieves state-of-the-art performance on outdoor driving datasets, with single-frame inference in seconds. It delivers high-fidelity 4D reconstructions, strong scalability, and superior cross-scene generalization—without scene-specific adaptation.

1 citationsRead paper