Fan Fei (费凡)费凡 (Fan Fei)

I am a fifth-year Ph.D. candidate at the Camera Intelligence Research Lab at Peking University, where I work on 3D computer vision and embodied intelligence. I am advised by Prof. Boxin Shi.

我是北京大学相机智能实验室的五年级博士生， 主要研究三维计算机视觉与具身智能，师从施柏鑫教授。

I received my B.S. (summa cum laude) in Intelligence Science and Technology from the Turing Class at the School of EECS, Peking University in 2022.

我于 2022 年在北京大学信息科学技术学院（School of EECS）图灵班 获得智能科学与技术专业荣誉学士学位。

I interned at Light Illusions under the supervision of Prof. Ping Tan from October 2023 to September 2024, and at the PKU-AI² Robotics Joint Lab of Embodied AI from June 2025 to May 2026, where I worked closely with Jiaming Liu and other researchers.

2023 年 10 月至 2024 年 9 月，我在 光影焕像 实习， 导师为谭平教授； 2025 年 6 月至 2026 年 5 月，我在北京大学-智平方具身智能联合实验室实习， 期间与刘家铭及其他研究者密切合作。

Since June 2026, I have been working at Tencent Hunyuan 3D as a research intern on PBR texture generation.

自 2026 年 6 月起，我在腾讯混元 3D 担任研究实习生， 从事 PBR 纹理生成方向的研究。

(Last updated: Sept. 2026)

（最后更新：2026 年 9 月）

Email邮箱  /  GitHub  /  Google Scholar  /  LinkedIn

Research研究方向

My research interests: 3D computer vision, computer graphics, and embodied intelligence, including inverse rendering, 3D reconstruction, texture generation, active perception, and end-to-end policy learning.

我的研究方向为：三维计算机视觉、计算机图形学与具身智能， 包括逆向渲染、三维重建、纹理生成、主动感知与端到端策略学习。

Twin-DAgger: Synergizing Digital Twins and Human Corrections for Efficient Robot Manipulation

We propose to treat each real-world intervention as both a spatial probe of failure-prone regions and a temporal prior of critical deviation moments to guide synthetic data generation, matching conventional DAgger performance with only 15% of the human interventions.

提出把每一次真实世界的人工干预同时视为定位易失败空间区域的空间探针、以及标记关键偏离时刻的时间先验， 以此指导合成数据生成，仅用 15% 的人工干预即可达到传统 DAgger 的效果。

PacTure: Efficient PBR Texture Generation on Packed Views with Visual Autoregressive Models

We propose a method to efficiently generate physically-based rendering (PBR) textures on a packed multi-view grid using a visual autoregressive model as multi-view generation backbone.

提出以视觉自回归模型作为多视角生成骨干，在打包的多视角网格上高效生成基于物理渲染（PBR）的材质贴图。

No Redundancy, No Stall: Lightweight Streaming 3D Gaussian Splatting for Real-time Rendering

We propose an algorithm/hardware co-design framework for lightweight streaming 3D rendering which achieves 5.41× speedup over the edge GPU baseline on average and up to 17.3× speedup with the customized accelerator.

提出一套面向轻量级流式三维渲染的算法/硬件协同设计框架，相比边缘 GPU 基线平均加速 5.41×，配合定制加速器最高可达 17.3× 加速。

SpecTRe-GS: Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting

We propose a 3D reconstruction method which models highly specular surfaces that reflect nearby objects through ray tracing in 3D Gaussian Splatting.

提出一种三维重建方法，通过在三维高斯泼溅中追踪光线，对反射邻近物体的高反光表面进行建模。

VMINer: Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources

We use near-field lights, such as flashlights, as sources of lighting variation to enhance both the practicality and the quality of multi-view inverse rendering.

利用手电筒等近场光源提供光照变化，从而提升多视角逆向渲染的实用性与质量。

SPLiT: Single Portrait Lighting Estimation Via a Tetrad of Face Intrinsics

From a single portrait image, we estimate a tetrad of face intrinsics and uses spherically distributed components to estimate lighting.

从单张人像图像出发，估计一组人脸本征属性，并利用分布到球面上的光照相关本征分量估计场景光照。

Design and source code from Jon Barron's website

页面设计与源码来自 Jon Barron 的主页