WeiJia Dou (窦为嘉)

WeiJia Dou (窦为嘉)

B.Eng. @Tongji Univ.
Incoming Ph.D. @Fudan Univ.

Academic Homepage of WeiJia Dou. Incoming Ph.D. student at Fudan University. Research interests: geometry-grounded visual foundation models, structured world models, and persistent video generation.

Geometry-Grounded Vision · Visual World Models · Persistent Video Generation

Grounding visual intelligence in the structure of the world, from open-vocabulary 3D segmentation to physically consistent and persistent video generation.

Geometry-Grounded Foundation Models Structured World Models Object-Centric Memory
EDU
B.Eng. in Automation at Tongji University (Class of 2026).
PHD
Incoming Direct Ph.D. student at Fudan University (Fall 2026), advised by Prof. Siyu Zhu.
FOCUS
Research: Structured visual representation learning for 3D perception and video generation, focusing on geometric consistency, object persistence, and long-horizon coherence.
OPEN
Open to collaboration on structured world models and 3D-consistent video generation; please reach out for citation or joint research.
Read Featured Paper Contact by Email Google Scholar

Research Updates

Recent milestones relevant for citation and collaboration.

  • 2026.07: SGC was accepted to ECCV 2026.
  • 2026.01: GeoPurify was accepted to ICLR 2026.
  • 2025.09: Admitted to the Direct Ph.D. program at Fudan University.

Selected Publication

Primary evidence block. If you plan to cite or collaborate, start here.

ECCV 2026
SGC
Measuring 3D Spatial Geometric Consistency in Dynamic Video Generation
Weijia Dou1, Wenzhao Zheng1, Weiliang Chen, Yu Zheng, Jie Zhou, Jiwen Lu✉
European Conference on Computer Vision (ECCV), 2026
arXiv Code
  • Introduced SGC, a diagnostic metric for evaluating 3D spatial geometric consistency in generated videos.
  • Disentangled static background stability from foreground dynamics via motion segmentation, depth estimation, and local camera pose divergence.
  • Revealed critical 3D geometric failures missed by FVD, VBench, and 2D consistency metrics across 1,296 real and generated videos.
ICLR 2026
GeoPurify
GeoPurify: A Data-Efficient Geometric Distillation Framework for Open-Vocabulary 3D Segmentation
Weijia Dou, Xu Zhang, Yi Bin✉, Jian Liu, Bo Peng, Guoqing Wang, Yang Yang, Heng Tao Shen
The 14th International Conference on Learning Representations (ICLR), 2026
arXiv OpenReview Code
  • Diagnosed feature inconsistency when projecting 2D VLM features into 3D.
  • Proposed GeoPurify with geometric contrastive distillation to recover latent 3D structure.
  • Reached SOTA with only ~1.5% annotated data.

Research Experience

2025.02 - 2025.09
Research Intern at Tsinghua University

Advised by Prof. Jiwen Lu and mentored by Dr. Wenzhao Zheng.

Topic 3D Spatial Geometric Consistency in Video Generation

Research Highlights
  • Showed that fidelity-centric metrics (e.g., FVD) miss geometric distortions.
  • Proposed SGC to quantify geometric instability via local camera-pose divergence.
2024.09 - Present
Research Intern at Tongji University

Advised by Prof. Yi Bin.

Topic Open-Vocabulary 3D Scene Understanding

Research Highlights
  • Analyzed the geometric gap in VLM-based 3D segmentation.
  • Developed GeoPurify to align 2D semantics with 3D structural priors.

Honors and Awards

Scholarships

  • 2025 Xiaomi Scholarship (Awarded to excellent students in technology fields)
  • 2024 National Scholarship (Highest Honor for Undergraduates in China, Top 0.2%)
  • 2024 Tongji Qidi Scholarship (Top 20 students in the college)
  • 2023 - 2025 Merit Student of Tongji University (Awarded for 3 consecutive years)

Competitions

  • 2024 "Star" Level Award (First Prize), The 18th "Challenge Cup" National Undergraduate Curricular Academic Science and Technology Works Competition ("Black Technology" Track).
  • 2024 Second Prize, The 7th National Undergraduate Embedded Chip and System Design Competition (FPGA Track).
  • 2024 Most Popular Project, The 17th National Conference on Undergraduate Innovation and Entrepreneurship.