Haichao Zhu

Senior Research Engineer · Computer Vision / 3D / Machine Learning

I work on vision and embodied intelligence systems that connect geometry, learning, and action, with growing interest in world representation and the intersection of 3D, LLMs, and agents.

GitHub · Google Scholar · Linkedin ·

Email: lszhuhaichao@gmail.com

About

I build computer vision and machine learning systems for problems that require both geometric reasoning and practical robustness. My background is in 3D reconstruction, visual-inertial perception, depth and pose estimation, and AR interaction.

More recently, I have been moving toward broader questions of world representation: how visual geometry, learned scene understanding, and agentic decision-making can work together in systems that perceive, reason, and act.

I am especially interested in projects that connect 3D perception with LLMs, multimodal reasoning, and interactive agents, while still grounding them in real data, real environments, and deployable systems.

Current Focus

  • Image-based gravity estimation for VIO/SLAM
  • 3D perception, reconstruction, and calibration
  • World representation for embodied and interactive systems
  • LLM- and agent-driven systems grounded in 3D understanding

Image Based Gravity Estimation for VIO/SLAM

A current research project on calibrating noisy IMU gravity priors from a single RGB image for downstream VIO, SLAM, and visual localization systems.

360° SLAM

A panoramic SLAM system for mapping and localization using 360-degree imagery with robust loop closure and global consistency.

Large-scale Indoor 3D Reconstruction from 360° Images

Learning- and geometry-driven reconstruction for large indoor spaces built from panoramic imagery.

Selected Papers

ARSketch

Sketch-based interaction for augmented reality glasses, published at ACM Multimedia 2020.

Deep Video Deinterlacing

TensorFlow implementation and research project based on real-time deep video deinterlacing.

HandAugment

Depth-based 3D hand pose estimation with two-stage hand region extraction and MANO-based data augmentation.

Globally Optimal Toon Tracking

SIGGRAPH 2016 work on globally optimized region correspondence and tracking for hand-drawn animation.

Contact

The easiest way to reach me is by email: lszhuhaichao@gmail.com