Wenzheng Chen

I'm a tenure-track Assistant Professor at Wangxuan Institute of Computer Technology, Peking University, where I mainly work on Computational Photography and 3D Vision. I am also affiliated in the Visual Computing and Learning Lab, collaborating with Prof. Baoquan Chen, Prof. Libin Liu, Prof. Mengyu Chu, and Dr. Bin Wang.

Before joining Peking University, I was a research scientist at NVIDIA Toronto AI Lab. I earned my Ph.D. from the University of Toronto and received both my Master's and Bachelor's degrees from Shandong University.

I am always actively recruiting Ph.D. students and research interns! Feel free to drop me a line with your CV and research statement!

Email  /  CV  /  Google Scholar  /  LinkedIn

Generally, my research focuses on computational photography and 3D vision. More specifically, I mainly explore how to utilize various imaging systems, including digital camera, LiDAR, structured light, and SPAD, etc, to perceive the world and predict its corresponding 3D attributes, such as 3D geometry, texture, surface material, environment light, and more. I am particularly interested in how to differentiate imaging systems and embed them within deep learning frameworks, which allows us to leverage all kinds of optical and physical prior knowledge inside imaging systems as well as the power of deep learning methods to boost 3D perception performance.

Thanks to my wonderful collaborators, my research has received several honors. My work on differentiable rendering was showcased at NVIDIA GTC (Bilibili, Youtube), presented by CEO Jensen Huang (黄仁勋), and has successfully turned into an Omniverse product. The popular YouTube technology channel Two Minute Papers has featured my work multiple times and collected millions of views(DIB-R, GET3D, NVdiffrec). My structured light work won the ICCP 2021 Best Poster Award. My NLOS work was exhibited at Princeton Art of Science Exhibition.

Full publication list in Google Scholar.

TurboSL provides sub-pixel-accurate surfaces and normals at mega-pixel resolution from structured light images, captured at fractions of a second.

EMDiffusion learns a clean diffusion model from corrupted data.

Combined with other NVIDIA technology, FEGR is one component of Neural Reconstruction Engine announced in GTC Sept 2022 Keynote.

We develop a 3D generative model to generate meshes with textures, bridging the success in the differentiable surface modeling, differentiable rendering and 2D GANs.

We propose a hybrid lighting representation to represent spatial-varying lighting for complex outdoor street scenes.

Nvdiffrec reconstructs 3D mesh with materials from multi-view images by combining diff surface modeling with diff renderer. The method supports Nvidia Neural Drivesim

DIB-R++ is a high-performant differentiable renderer which combines rasterization and ray-tracing together and supports advanced lighitng and material effects. We further embed it in deep learning and jointly predict geometry, texture, light and material from a single image.

We explore StyleGAN as a multi-view image generator and train inverse graphics from StyleGAN images. Once trained, the invere graphics model further helps disentangle and manipulate StyleGAN latent code from graphics knowledge. Our work was featured at NVIDIA GTC 2021 and has become an Omniverse product.

We propose to learn feature embeddings for non-line-of-sight imaging and recognition by propagating features through physical modules.

We predict deformable tetrahedral meshes from images or point clouds, which support arbitrary topologies. We also design a differentiable renderer for tetrahedron, allowing 3D reconstrucion from 2D supervison only.

We present optical SGD, a computational imaging technique that allows an active depth imaging system to automatically discover optimal illuminations & decoding.

An interpolation-based 3D mesh differentiable renderer that supports vertex, vertex color, multiple lighting models, texture mapping and could be easily embedded in neural networks.

We show hidden objects can be recovereed from conventional images instead of transient images.

We predict object polygon contours from graph neural networks, where a novel 2D differentiable rendering loss is introduced. It renders a polygon countour into a segmentation mask and back propagates the loss to help optimize the polygon vertices.

alacarte designs structured light patterns from a maching learning persepctive, where patterns are automatically optimized by minimizing the disparity error under any given imaging condition.

3D pose estimation from model trained with synthetic data and domain adaptation.

I stole the website template from Jon Barron.