Selected work

Representative Papers

GSMap: 2D Gaussians for Online HD Mapping

May 10, 2026

Existing high-definition map construction methods struggle to simultaneously achieve geometric accuracy and topological correctness: vectorization-based approaches preserve structural integrity but suffer from geometric distortions, whereas rasterization-based methods offer precise geometry yet lack explicit structural representation. To address this limitation, this work proposes GSMap, a novel framework that introduces learnable 2D Gaussian sequences to represent map elements, modeling vector vertices as Gaussian centers. By integrating differentiable rasterization for pixel-level geometric constraints and topology-aware vectorization to enforce structural regularity, GSMap enables end-to-end joint optimization of geometry and topology. The method significantly outperforms existing approaches on both nuScenes and Argoverse2 benchmarks while remaining compatible with mainstream HD map architectures.

0 citationsRead paper

LiDAR Prompted Spatio-Temporal Multi-View Stereo for Autonomous Driving

Mar 04, 2026

Existing deep learning–based depth estimation methods for autonomous driving struggle to simultaneously achieve high metric accuracy, multi-view and temporal consistency, and cross-domain generalization. This work proposes DriveMVS, a novel framework that uniquely integrates sparse yet precise LiDAR measurements as both hard geometric priors and soft feature guidance within a multi-view stereo matching network. Furthermore, a spatiotemporal consistency decoder is introduced to fuse cues from multiple views, temporal sequences, and LiDAR observations into an optimized cost volume. Extensive experiments demonstrate that DriveMVS significantly outperforms state-of-the-art methods across multiple benchmarks, achieving leading performance in metric accuracy, inter-frame stability, and zero-shot cross-domain transferability.

0 citationsRead paper

TIGaussian: Disentangle Gaussians for Spatial-Awared Text-Image-3D Alignment

Jan 27, 2026

This work addresses the challenges of 3D feature extraction and cross-modal alignment among text, images, and 3D data by proposing a spatially aware multimodal alignment framework. The method introduces a novel multi-branch tokenizer that decouples the 3D Gaussian Splatting (3DGS) representation into compact latent codes, effectively integrating multi-view features with diffusion priors to mitigate view ambiguity. Furthermore, a text-to-3D adaptive projection module is designed to enable fine-grained cross-modal alignment. Extensive experiments demonstrate that the proposed approach achieves state-of-the-art performance across multiple benchmark datasets in tasks including cross-modal retrieval, zero-shot classification, and scene recognition.

0 citationsRead paper

Replay Failures as Successes: Sample-Efficient Reinforcement Learning for Instruction Following

Dec 29, 2025

Large language models (LLMs) often produce responses that partially satisfy instruction constraints, leading to sparse reward signals and low sample efficiency in instruction-following reinforcement learning. Method: We propose Hindsight Instruction Replay (HiR), a novel framework that (i) retrospectively reconstructs failed responses by grouping them according to satisfied constraints to generate high-quality pseudo-positive samples; (ii) formulates dual-granularity preference learning objectives—both instruction-level and response-level; and (iii) performs efficient optimization using only binary reward signals. HiR integrates constraint-aware filtering-and-rewriting, reinforcement-learning-driven sample replay, and dual preference modeling. Results: HiR achieves significant performance gains across diverse complex instruction-following tasks. It attains comparable or superior results with over 30% reduction in sampling and computational overhead. We publicly release our code and dataset to foster reproducibility and further research.

0 citationsRead paper

GUIDE: Gaussian Unified Instance Detection for Enhanced Obstacle Perception in Autonomous Driving

Nov 16, 2025

To address the challenge of representing irregular obstacles with 3D bounding boxes in autonomous driving, this paper proposes GUIDE—a unified framework leveraging 3D Gaussian ellipsoids for both instance detection and occupancy prediction. Methodologically, we introduce a sparse Gaussian-to-voxel projection strategy that preserves accuracy while substantially reducing computational overhead; further, we integrate instance-level occupancy prediction with multi-frame Gaussian parameter association to achieve robust 3D instance tracking. On the nuScenes benchmark, GUIDE achieves an instance occupancy mAP of 21.61—50% higher than the state-of-the-art—while also delivering significant gains in detection and tracking performance. Key contributions include: (1) the first unified modeling of 3D Gaussians across perception tasks; (2) a novel sparse Gaussian–voxel projection mechanism; and (3) an end-to-end trainable joint framework for instance occupancy prediction and tracking.

0 citationsRead paper

Recent publications

Latest Papers

GSMap: 2D Gaussians for Online HD Mapping

May 10, 2026

Existing high-definition map construction methods struggle to simultaneously achieve geometric accuracy and topological correctness: vectorization-based approaches preserve structural integrity but suffer from geometric distortions, whereas rasterization-based methods offer precise geometry yet lack explicit structural representation. To address this limitation, this work proposes GSMap, a novel framework that introduces learnable 2D Gaussian sequences to represent map elements, modeling vector vertices as Gaussian centers. By integrating differentiable rasterization for pixel-level geometric constraints and topology-aware vectorization to enforce structural regularity, GSMap enables end-to-end joint optimization of geometry and topology. The method significantly outperforms existing approaches on both nuScenes and Argoverse2 benchmarks while remaining compatible with mainstream HD map architectures.

0 citationsRead paper

LiDAR Prompted Spatio-Temporal Multi-View Stereo for Autonomous Driving

Mar 04, 2026

Existing deep learning–based depth estimation methods for autonomous driving struggle to simultaneously achieve high metric accuracy, multi-view and temporal consistency, and cross-domain generalization. This work proposes DriveMVS, a novel framework that uniquely integrates sparse yet precise LiDAR measurements as both hard geometric priors and soft feature guidance within a multi-view stereo matching network. Furthermore, a spatiotemporal consistency decoder is introduced to fuse cues from multiple views, temporal sequences, and LiDAR observations into an optimized cost volume. Extensive experiments demonstrate that DriveMVS significantly outperforms state-of-the-art methods across multiple benchmarks, achieving leading performance in metric accuracy, inter-frame stability, and zero-shot cross-domain transferability.

0 citationsRead paper

TIGaussian: Disentangle Gaussians for Spatial-Awared Text-Image-3D Alignment

Jan 27, 2026

This work addresses the challenges of 3D feature extraction and cross-modal alignment among text, images, and 3D data by proposing a spatially aware multimodal alignment framework. The method introduces a novel multi-branch tokenizer that decouples the 3D Gaussian Splatting (3DGS) representation into compact latent codes, effectively integrating multi-view features with diffusion priors to mitigate view ambiguity. Furthermore, a text-to-3D adaptive projection module is designed to enable fine-grained cross-modal alignment. Extensive experiments demonstrate that the proposed approach achieves state-of-the-art performance across multiple benchmark datasets in tasks including cross-modal retrieval, zero-shot classification, and scene recognition.

0 citationsRead paper

Replay Failures as Successes: Sample-Efficient Reinforcement Learning for Instruction Following

Dec 29, 2025

Large language models (LLMs) often produce responses that partially satisfy instruction constraints, leading to sparse reward signals and low sample efficiency in instruction-following reinforcement learning. Method: We propose Hindsight Instruction Replay (HiR), a novel framework that (i) retrospectively reconstructs failed responses by grouping them according to satisfied constraints to generate high-quality pseudo-positive samples; (ii) formulates dual-granularity preference learning objectives—both instruction-level and response-level; and (iii) performs efficient optimization using only binary reward signals. HiR integrates constraint-aware filtering-and-rewriting, reinforcement-learning-driven sample replay, and dual preference modeling. Results: HiR achieves significant performance gains across diverse complex instruction-following tasks. It attains comparable or superior results with over 30% reduction in sampling and computational overhead. We publicly release our code and dataset to foster reproducibility and further research.

0 citationsRead paper

GUIDE: Gaussian Unified Instance Detection for Enhanced Obstacle Perception in Autonomous Driving

Nov 16, 2025

To address the challenge of representing irregular obstacles with 3D bounding boxes in autonomous driving, this paper proposes GUIDE—a unified framework leveraging 3D Gaussian ellipsoids for both instance detection and occupancy prediction. Methodologically, we introduce a sparse Gaussian-to-voxel projection strategy that preserves accuracy while substantially reducing computational overhead; further, we integrate instance-level occupancy prediction with multi-frame Gaussian parameter association to achieve robust 3D instance tracking. On the nuScenes benchmark, GUIDE achieves an instance occupancy mAP of 21.61—50% higher than the state-of-the-art—while also delivering significant gains in detection and tracking performance. Key contributions include: (1) the first unified modeling of 3D Gaussians across perception tasks; (2) a novel sparse Gaussian–voxel projection mechanism; and (3) an end-to-end trainable joint framework for instance occupancy prediction and tracking.

0 citationsRead paper