计算机视觉算法-暑期实习生（三维数字人方向）

优缺点分析

角色解读

• Build a strong publication record in top-tier conferences (CVPR, ICCV, etc.) with potential for full-time research role.
• Develop expertise in 3D human modeling and generative AI, leading to senior researcher or tech lead positions.
• Transition to product-focused roles, applying research to real-world beauty imaging products at Meitu.
• Research and implement algorithms for 3D human reconstruction and tracking, including depth estimation and multi-view generation.
• Reproduce and improve state-of-the-art papers in 3D vision and generative models.
• Build tools for data processing, evaluation, and visualization to accelerate algorithm development.
• Collaborate with the data team to enhance the 3D beautification pipeline for high-quality scaling.
• Strong Python programming and deep learning (PyTorch) skills, with experience in libraries like numpy, transformer, open3d.
• Solid understanding of computer vision fundamentals, especially 3D reconstruction, human pose estimation, and geometry.
• Ability to read and reproduce research papers independently.
• Familiarity with diffusion models, autoregressive models, or flow matching is a plus.

申请策略

• Demonstrate proactive learning and ability to tackle challenging problems in your cover letter.
• Highlight alignment with Meitu's focus on aesthetics and user experience.
• Emphasize any prior research or projects in 3D reconstruction, human pose estimation, or SMPL.
• Showcase proficiency in Python and PyTorch with links to GitHub repositories.
• Mention any publications, open-source contributions, or relevant competitions.
• Practice reproducing recent papers in 3D vision and generative models (e.g., diffusion models).
• Deepen understanding of SMPL model, camera geometry, and multi-view geometry.

面试指南

• Use the STAR method (Situation, Task, Action, Result) for experience-based questions.
• For technical questions, outline the problem, propose a solution with key steps, discuss trade-offs, and suggest evaluation metrics.
• Explain how you would approach 3D human reconstruction from a single image.
• Describe a time you reproduced a research paper and any challenges faced.
• How would you fuse SMPL with depth estimates for robust tracking?
• What are the advantages and limitations of diffusion models for image generation?
• How do you evaluate the quality of 3D reconstructions?
• Review recent papers from CVPR, ICCV on human reconstruction and generative models.