跨越维度 迈向未来
跨维智能是由顶尖学术、工程团队共同组建的具身智能解决方案公司。自成立以来,跨维智能专注于运用前沿科研成果,构建DexVerse™ ,AI 和合成数据引擎,以形成通用机器人智能。 跨维智能将前沿技术运用在了工业场景,彻底改变了工业自动化生产,为具身智能技术的规模化应用做出了开创性的贡献。
特色研究成果
以生成式AI 2.0实现规模化仿真与多模态数据生成
Fantasia3D: Disentangling Geometry and Appearance for High-quality Text-to-3D Content Creation
ComboVerse: Compositional 3D Assets Creation Using Spatially-Aware Diffusion Guidance
arXiv preprint arXiv:2403.12409, 2024.
Sam-6d: Segment anything model meets zero-shot 6d object pose estimation
3D多模态感知及机器人操作大模型
Open-set semi-supervised learning for 3d point cloud understanding
Submitted to SIGGRAPH ASIA,2024.
Grasp Proposal Networks: An End-to-End Solution for Visual Learningof Robotic Grasps
Neural Information Processing Systems (NeurlPS),2020.
3D AffordanceNet: A Benchmarkfor Visual Object
lEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2021
创新成果有效应对Sim2Real数据与真实数据的域差别
A New Benchmark: On the Utilityof Synthetic Data for BareSupervised Learning and Downstream Domain Adaptation
lEEE Conference on ComputerVision and Pattern Recognition(CVPR), 2023.
Quasi-Balanced Self-Training onNoise-Aware Synthesis of 0bjectPoint Clouds for Closing Domain Gap
European Conference onComputer Vision (ECCV), 2022.
Analytic-Splatting: Anti-Aliased 3DGaussian Splatting via Analytic Integration
arXiv:2403.11056, 2024.
Real2Sim支持低成本获取高质量 3D 数字资产
VISTA: Boosting 3D Object Detection via Dual Cross-VIew SpaTial Attention
lEEE Conference on ComputerVision and Pattern Recognition(CVPR), 2023.
A Deep Learning Based Interactive Sketching System for Fashion Images Design
European Conference onComputer Vision (ECCV), 2022.
TANGO: Text-driven PhotoreAlistic aNd Robust 3D Stylization via LiGhting DecompOsition
CVPR, 2024.
Sparse Steerable Convolutions: An Efficient Learning of SE(3)-Equivariant Features for Estimation and Tracking of Object Poses in 3D Space
arXiv:2403.11056, 2024.
颠覆性原创成果 荣获多项冠军
- 荣获多项竞赛冠军
- 参加IROS物体姿态估计竞赛,获多项冠亚军,参赛队伍包括商汤、清华等团队
- 研发F-ConvNet, HotSpotNet等3D检测模型,在KITTI等国际无人驾驶基准排行榜(Pedestrian检测)长期保持第一, 力压华为等企业机构的方案
- 研发DualPoseNet 全自由度姿态估计模型,精确度在NOCS基准数据库保持第一,大幅领先其他方案
- 研发SRDC域适应学习模型, 在GTA5->Cityscapes和SYNTHIA->Cityscapes等合成数据到真实场景的无人驾驶迁移学习基准数据库上精确度保持第一
- 参加IEEE BTAS 2016 Video Person Recognition Evaluation挑战赛,在国内外众多参赛队伍中排名第一
- 多项颠覆性原创成果
- 研发智能视觉碎石系统,应用于芬兰采石场,提升碎石采石效率
- 研发GPNet等模型,实现深度学习无注册物体抓取,性能超过Nvidia所研发模型,公布最大仿真无注册抓取数据集
- 全球首创3D AffordanceNet三维功能可供性分析方法及大规模基准数据集,助力学术及工业界人机交互研发
- 研发SkeletonNet,ToMoNet等网络,首次实现深度学习复杂拓扑表面生成,入选人工智能顶会CVPR 19最佳论文候选
- 首创Analytic Marching无损解析表面网格理论及算法,开源AnalyticMesh软件,有望取代Marching Cubes成为网格提取与存储行业新标准
- 提出深度模型自适应优化方法,打通传统MVS技术与深度表面重建壁垒,大幅提升MVS重建效果