Yuqing Lan

I am a PhD candidate at NUDT, where I am advised by Professor Kai Xu, Professor Chenyang Zhu and Professor Yijie Wang. My PhD research focuses on 3D vision, embodied intelligence, vision foundation models, and 3D scene understanding. I am particularly interested in enabling machines to perceive, represent, and reason about complex real-world environments from visual observations. My works aim to bridge geometric perception and semantic understanding for more robust and generalizable embodied systems.

Email  /  Github

Research

I'm interested in 3D vision, embodied intelligence, and vision-language models. My research focuses on enabling machines to understand, represent, and reason about the physical world from visual observations.

RemixFusion is a residual-based mixed representation for scene reconstruction and camera pose estimation, dedicated to high-quality and large-scale online RGB-D reconstruction.

BoxFusion is a reconstruction-free online framework for open-vocabulary 3D object detection that fuses 2D detections from a pre-trained visual fundation model across multiple views into unified 3D boxes via correspondence matching and IoU-guided optimization. It enables real-time performance and achieves SOTA without needing expensive dense 3D reconstruction.

To achieve online 3D open-vocabulary segmentation during real-time scene reconstruction, we propose a fast method of 2D masks lifting by using voxel hashing for efficient 3D scene querying.

We propose to mine object functionality with user preference alignment directly from the scene itself through LLM-enhanced scene graph learning which transforms the input scene graph into an affordance-enhanced graph with information-enhanced nodes and newly discovered edges.

ARM3D is a 3D object detection model that exploits contextual relationships to enhance 3D object understanding.

The core idea of DisARM is that contextual information is critical to tell the difference between different objects when the instance geometry is incomplete or featureless. We find that relations between proposals provide a good representation to describe the context.

3DRM is a relation module that improves the accuracy and robustness of 3D object detection by explicitly modeling the semantic and spatial relationships between objects for enhanced contextual reasoning.

Miscellanea