Yizhi Wang (王逸之)

I am a Postdoc Fellow working with Professor Hao (Richard) Zhang in GrUVi Lab at Simon Fraser University (SFU), Canada. I am now working on 3D Shape Reconstruction and Content Creation.

I received my PhD degree from Peking University, where I worked on Computer Vision and Computer Graphics. During my PhD, I was focused on applying deep generative models to analyze and synthesize 2D geometric data (e.g., glyph, font, and layout). I was supervised by Prof. Zhouhui Lian and Prof. Jianguo Xiao.

Email  /  CV  /  Google Scholar  /  Github  /  Twitter

• 07/2024 SweepNet is accepted by ECCV 2024.
• 05/2024 EASI-Tex is accepted by SIGGRAPH 2024 as a Journal Paper.
• 03/2024 Slice3D is accepted by CVPR 2024. Code has been released. Demo is coming out!
• 01/2024 I am co-organzing the Workshop on Graphic Design Understanding and Generation in CVPR 2024.
• 12/2023 Introducing Slice3D, a new notion for single-view 3D reconstruction via multi-slice reasoning.
• 07/2023 DS-Fusion is accepted by ICCV 2023. Check out the fun video!
• 06/2023 Check how Christina uses DeepVecFont in her art-design project (Video and Blog)!
• 05/2023 Introducing BRICS, a new feature representation method, which can generate images with more accurate structures and less distortions.

I have strong interests in shape synthesis and analysis. Specifically, my researching projects cover the following topics: 3D Reconstruction, Font Generation, Layout Generation, Scene Text (Character) Recognition and Detection, Font Recognition, etc. Representative papers are highlighted.

Our single-view 3D reconstruction method, Slice3D, predicts multi-slice images to reveal occluded parts without changing the camera (in contrast to multi-view synthesis), and then lifts the slices into a 3D model.

A novel approach for shape abstraction via sweep surfaces, which utilize superellipses for profile representation and B-spline curves for the axis.

A novel approach for single-image mesh texturing, which employs diffusion models with judicious conditioning to seamlessly transfer textures.

A novel shape encoding for learning neural field representation of shapes that is category-agnostic and generalizable amid significant shape variations.

Automatically generate an artistic typography by stylizing one or more letter fonts to visually convey the semantics of an input word.

A bi-levelfeature representation for an imagecollection, consisting of a per-image latent space on top of a multi-scale feature grid space.

Employing Transformers and a relaxed representation for higher-quality vector font generation (support Chinese glyphs).

A content-aware layout generation network which takes element images and their corresponding content (such as texts) as input and synthesizes aesthetic layouts for them automatically.

Directly synthesize vector fonts via dual-modality learning and differentiable rasterization (rendering), instead of vectorizing synthesized glyph images by rule-based methods.

Automatically synthesize fonts according to user-specified attributes (such as italic, serif, cursive, and angularity) and their corresponding values.

Address the challenge of font variance in STR and propose a font-independent feature representation method to increase the robustness of STR models.

Accurately recognize the font styles of texts in natural images by proposing a novel method based on deep learning and transfer learning.

• 07/2022 I was invited to give a talk (in Chinese) by CSIG (China Society of Image and Graphics).

template adapted from Songyou and Jon.