Yuliang Guo

I'm a Lead (Staff) Research Scientist at Bosch AI Research Center at Silicon Valley, where I lead a team that works on 3D vision and interactive AI solutions.

At Bosch, my work focuses on pioneering computer vision and AI research, aimed at enabling generalization across diverse hardware configurations and embodiments. My research has been integrated into several Bosch products or prototypes, such as assisted driving technologies, industrial augmented reality (AR) systems, and AI-powered indoor robotics.

Email  /  CV  /  Bio  /  Scholar  /  Github

Research

My recent research interests are centered on formulating active 3D perception methods adaptable across various embodiments, as well reconstructing interactable scene models from real world. My work aims to foster intelligent interactions between humans and AI while upholding safety in the physical world.

A monocular object reconstruction framework effectively integrating object pose estimation and NeRF-based reconstruction. A novel camera-invariant pose estimation module is introduced to resolve depth-scale ambiguity and enhance cross-domain generalization.

An advanced Gaussian Splatting method effectively fusing Lidar and surrounding camera views for autonomous driving. The method uniquely leverages an intermediate occ-tree feature volume before GS such that GS parameters can be initialized from feature-volume-generated 3D surface more effectively.

An effective framework leveraging lightweight and scalable priors-Standard Definition (SD) maps in the estimation of online vectorized HD map representations.

A mathematical framework to prove that the dice loss leads to superior noise-robustness and model convergence for large objects compared to regression losses. A flexible monocular 3D detection pipeline integrated with bird-eye view segmentation.

The first neural reconstruction method able to complete the occluded surfaces from large scenes. A key enabler to build interactable environments from real world, generalizing robotic reinforcement learning via reduced domain gap.

A physically plausible indoor 3D object insertion approach to automatically "copy" virtual objects and "paste" them into real scenes.

A keypoint-based object-level SLAM framework that can provide globally consistent 6DoF pose estimates for symmetric and asymmetric objects.

The first transformer approach to handle 360 monocular depth estimation with spherical distortion. Novel designs include tangent-image coordinate embedding and geometry-aware feature fusion.

A real-time method to predict multi-person 3D poses from a depth image. Introduce new part-level representation to enables an explicit fusion process of bottom-up part detection and global pose detection. A new 3D human posture dataset with challenging multi-person occlusion.

A pioneer work in predicting 3D lanes from a single image with high generalization to novel scenes. A 3D lane synthetic dataset is introduced.

Numerically robust techniques to precisely estimate differential geometry attributes associated with image edges, including localization, orientation, and curvature, as well as edge topology. A curve fragment dataset is introduced for the evaluation of precise geometric attributes.

A novel approach for estimating the 2D pose of an articulated object with an application to automated video analysis of small laboratory animals.

One-shot learning gesture recognition on RGB-D data recorded from Microsoft Kinect. A novel bag of manifold words (BoMW) based feature representation on sysmetric positive definite (SPD) manifolds.

This study investigates the relative diagnosticity and the optimal combination of multiple cues (we consider luminance, color, motion and binocular disparity) for boundary detection in natural scenes. A multi-cue boundary dataset is introduced to facilitate the study.

A multi-stage approach to curve extraction where the curve fragment search space is iteratively reduced by removing unlikely candidates using geometric constrains, but without affecting recall, to a point where the application of an objective functional becomes appropriate.

Selected Patents