CVPR 24
Posts by Collection
portfolio
CVPR 24
西湖大学 于开丞老师
达摩院
硕凡导师 
publications
BossNAS Family: Block-wisely Self-supervised Neural Architecture Search
Published in , 1900
Download here
Exploring Hybrid CNN-transformers with Block-wisely Self-supervised Neural Architecture Search
Published in , 1900
Download here
BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework
Published in , 1900
Download here
Bevheight++: Toward robust visual centric 3d object detection
Published in , 1900
Download here
BEVHeight: A Robust Framework for Vision-based Roadside 3D Object Detection
Published in , 1900
Download here
AlignMiF: Geometry-Aligned Multimodal Implicit Field for LiDAR-Camera Joint Synthesis
Published in , 1900
Download here
MLP Can Be A Good Transformer Learner
Published in , 1900
Download here
LiDAR-NeRF: Novel LiDAR View Synthesis via Neural Radiance Fields
Published in , 1900
Download here
Opensight: A simple open-vocabulary framework for lidar-based object detection
Published in , 1900
Download here
Making large language models better planners with reasoning-decision alignment
Published in , 1900
Download here
LiT: Unifying LiDAR Languages with LiDAR Translator
Published in , 1900
Download here
BEV-TSR: Text-Scene Retrieval in BEV Space for Autonomous Driving
Published in , 1900
Download here
UniGS: Unified Language-Image-3D Pretraining with Gaussian Splatting
Published in , 1900
Download here
S2-Track: A Simple yet Strong Approach for End-to-End 3D Multi-Object Tracking
Published in , 1900
Download here
RoboPearls: Editable Video Simulation for Robot Manipulation
Published in , 1900
Download here
OmniGen: Unified Multimodal Sensor Generation for Autonomous Driving
Published in , 1900
Download here
CorrectAD: A Self-Correcting Agentic System to Improve End-to-end Planning in Autonomous Driving
Published in , 1900
Download here
LiDAR-GS++: Improving LiDAR Gaussian Reconstruction via Diffusion Priors
Published in , 1900
Download here
talks
实习1: 暗物智能-自动机器学习实验室-算法实习生
Published:
神经网络结构搜索和自监督学习:1) 提出结构搜索方法 BossNAS，通过分块自监督机制缓解权重共享空间过大 和监督偏差问题，显著提升模型排序准确性与搜索效果。 2) 设计自动化自监督视图生成方法 AutoView，避免繁琐手工视图设计，以对抗方式最小化视图间互信息;提出自约束损失函数，联合优化互信息以提升自监督学习。
实习2: 阿里巴巴-达摩院-自动驾驶实验室-感知算法实习生
Published:
自动驾驶 多模态感知: 1) 主导构建 RobustBenchmark，收集 BEV 感知中常见噪声干扰样例，开发系统化鲁棒 性测试工具包;在 nuScenes 和 Waymo 上建立新鲁棒性基准，全面评估主流 BEV 融合方法的抗干扰性能。2) 参与 BEVFusion 框架设计:采用独立网络分别编码雷达与图像特征，统一投影至 BEV 空间再进行融合，实现近似 后融合的结构，打破视觉对雷达的主导依赖。 3) 参与设计 BEVHeight:为提升检测在参数扰动下的稳定性， 提出以目标高度回归替代直接深度预测，通过几何转换间接获取深度，显著缓解相机参数扰动对性能的影响。4) 参与开发 Opensight:提出基于 LiDAR 的开放词汇检测框架。通过图像生成通用 2D 检测框并回投至 LiDAR 空间估计 3D 位置;设计跨模态对齐与融合模块，将 3D 与 2D 特征对齐后进行语义解码，实现开放类目标检测能力。
实习3: 理想汽车-自动驾驶实验室-世界模型实习生
Published:
• 自动驾驶场景 多模态生成: 1) 参与设计 Delphi，一种基于扩散模型的长视频生成方法，引入跨视角共享噪声机 制与特征对齐模块以增强空间与时间一致性。 2) 提出首个端到端多模态场景生成框架 OmniGen，在统一 0-BEV 表征中融合 LiDAR 与图像，通过扩散模型生成多模态数据。 3) 提出 BEV-TSR，首个基于 BEV 空间的文本场 景检索框架，结合知识图谱与大语言模型增强语义理解，实现高精度多模态检索。