3D Reconstruction
Feed-forward and sparse-view reconstruction, multi-view geometry, cross-view reasoning, 3D Gaussian Splatting, and scalable 3D foundation models.
Ph.D. Student · 3D Vision and Generative Models
Feng Zhou (周峰)
I am a Ph.D. student in Control Science and Engineering at Beijing University of Posts and Telecommunications, advised by Prof. Jianqin Yin at the BUPT-COST Lab. I am expected to graduate in June 2027.
My research lies at the intersection of 3D vision and generative models, with an emphasis on feed-forward 3D reconstruction, 3D world models, and controllable visual generation. I received my B.Eng. degree in Internet of Things Engineering from BUPT in 2022.
I am currently seeking internship or full-time opportunities in 3D vision, world models, and generative AI.
News
- A paper on 3D vision was accepted to IEEE TVCG.
- A paper on 3D vision was accepted to SIGGRAPH Asia 2026.
- A paper on spatial audio was accepted to ICME 2026.
- A paper on generative models was accepted to CVPR 2026.
- A paper on generative models was accepted to AAAI 2026 as an oral presentation.
- A paper on generative models was accepted to IEEE TPAMI 2025.
- A paper on 3D vision was accepted to IEEE TCSVT 2025.
- A paper on generative models was accepted to CVPR 2025.
- A paper on 3D vision was accepted to AAAI 2024.
Research
3D World Models
Structured 3D scene representations and generation, including scene-level latent modeling, flow-based generation, and editable 3DGS environments.
Controllable Generation
Diffusion and flow models for controllability transfer, in-domain generation, and resolution extrapolation across U-Net- and Transformer-based architectures.
Experience
InSpatio
Spatial Intelligence and 3D World Model Intern
- Engineering: Contribute to TOPOS1.0 3D world model development, with responsibility for full-scene point-cloud alignment and 3DGS editing.
- Research: Study VAE representations and flow-matching training for native 3D scene generation.
Horizon Robotics
Special Talent Program Intern · 3D Reconstruction Foundation Model R&D
Worked on feed-forward 3D reconstruction foundation models, including sparse attention, selective cross-view communication, high-resolution reconstruction, large-scale distributed training, and multi-source data construction and evaluation.
Publications
* Equal contribution
Initialize to Generalize: A Stronger Initialization Pipeline for Sparse-View 3DGS
IEEE TVCG 2026
Builds a stronger sparse-view 3DGS initialization pipeline using frequency-aware SfM, 3DGS self-initialization, and point-cloud regularization.
GeoWeave: Learning Reliable Cross-View Dependencies for Feed-Forward 3D Reconstruction
SIGGRAPH Asia 2026
Introduces selective cross-view communication to improve pose and point-cloud reconstruction under weak overlap and distracting views.
DSD: Dual Self-Distillation for Robust Audio Tagging
ICME 2026
Improves robust audio tagging through hierarchical intra-layer supervision and EMA-based mask-consistency self-distillation.
ResDiT: Evoking the Intrinsic Resolution Scalability in Diffusion Transformers
CVPR 2026
Studies resolution extrapolation in Diffusion Transformers through positional encoding, attention receptive fields, and frequency-aware detail preservation.
Exploring Position Encoding in Diffusion U-Net for Training-free High-resolution Image Generation
AAAI 2026 · Oral Presentation
Reveals the implicit positional role of convolutional zero padding and introduces a training-free boundary-complement strategy for high-resolution diffusion inference.
Controllable Generation with Text-to-Image Diffusion Models: A Survey
IEEE TPAMI 2025
Surveys controllable text-to-image diffusion across condition injection, structural control, editing, and personalization.
OMEGAS: Object Mesh Extraction from Large Scenes Guided by Gaussian Segmentation
IEEE TCSVT 2025
Extracts object-level meshes from large 3D scenes using Gaussian segmentation and diffusion priors for occluded regions.
Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation
CVPR 2025
Transfers controllable generation capabilities to image-only target domains through guidance-decoupled prior preservation.
Lifting by Image - Leveraging Image Cues for Accurate 3D Human Pose Estimation
AAAI 2024
Uses pose-guided attention and adaptive feature selection to improve image-aware 2D-to-3D human pose lifting.
Research Journey & Vision
My research journey has unfolded in two stages.
The first began in the summer of 2022. During this period, I entered computer vision through concrete research problems and gradually developed my own understanding of deep learning. Rather than committing to a fixed direction, I focused on building a broad view of the field. Over time, I moved beyond individual models and began to think more systematically about how visual problems are formulated, how data and supervision shape what a model learns, and how architecture and optimization affect its generalization. This process gave me a stronger technical foundation and a broader perspective on how different ideas and learning paradigms relate to one another.
The second stage began in late 2025. By then, I had started to form my own view of how artificial intelligence was evolving, and my personal research direction gradually became clearer. I found a long-term goal that I hope to pursue throughout my career: contributing to intelligent systems capable of learning from and reasoning about the physical world, with the ultimate ambition of helping humanity deepen its understanding of nature and perhaps even uncover new physical laws.