CEO, 深圳市视元引擎有限公司.
Seeking research and industry collaborators in 3D vision, neural rendering, computer graphics, and avatar reconstruction.
I lead 深圳市视元引擎有限公司, developing low-cost 2D-to-3D conversion, 3D content generation, and hardware-interoperable glasses-free 3D display systems.
My research focuses on 3D/4D computer vision and geometry-aware generative modeling, with particular interests in depth estimation, novel-view synthesis, 3D Gaussian Splatting, human avatars, and evidence-preserving medical visualization.
I work across data preparation, geometry, PyTorch training, evaluation, GPU inference, and interactive visualization, with a focus on reliable monocular and sparse-view reconstruction under occlusion and limited observations.
Research question. How can missing satellite observations be reconstructed from temporal evidence without smoothing away genuine spatial change?
My contribution. I developed the EMAE temporal masked-autoencoder baseline and the MALA extension, combining temporal attention, multi-scale spatial features, explicit masks, and optional LaMA initialization.
Main result. Across four currently reported representative cases, MALA averages 31.05 dB PSNR, compared with 28.89 dB for EMAE. The related paper is under review, and the work produced two patent applications.
Research question. Can explicit depth reprojection and localized completion synthesize stable multi-view output more efficiently than regenerating every target view?
My contribution. I designed the depth-warping system, the hole/valid/pollute reliability representation, the lightweight MGMI completion network, and the synthetic defect-generation pipeline.
Main result. The research prototype reports approximately 5 ms per 4K target-view completion and approximately 15 FPS for 11-view output on one RTX 4090. Related work was accepted at ICDT 2026.
Research question. Can frozen relative-depth features predict per-frame scale and shift parameters for scene-conditioned metric-depth calibration?
My contribution. I adapted Video Depth Anything for single-image feature extraction, designed reciprocal-affine calibration, generated Depth Pro pseudo-labels, and trained a lightweight scale head while freezing the backbone.
Main result. The current 1,800-frame scene snapshot reports teacher-referenced MSE of 0.1294/0.2248 on two held-out cameras. These values measure in-scene teacher agreement, not cross-scene generalization.
Research question. Can relative monocular depth regularize underconstrained side-view facial geometry in RGB-only avatar reconstruction?
My contribution. I extended FlashAvatar with aligned depth loading, scale-invariant Pearson-correlation supervision, conservative face masks, preprocessing validation, and novel-view comparison tools.
Main result. The depth-supervision branch runs end to end and has produced qualitative cross-view comparisons on three identities. A controlled quantitative benchmark is not yet available.
Research question. How can MRI evidence, reviewed semantic structures, and local appearance representations be combined without making clinically unsupported claims?
My contribution. I built the privacy-minimized data schema and export tools, WebGL2 volume/cutaway viewer, segmentation interfaces, MRI-derived Gaussian pipeline, hybrid manifest, and automated tests.
Main result. The public prototype delivers one authorized privacy-minimized case with four MRI sequences and a 47,443-Gaussian representation. It is an engineering prototype, not a clinically validated system.
- DWvs: Depth-Guided Image Warping and Hole Filling for Novel View Synthesis. Co-author, ICDT 2026, accepted.
- Research on Spectral Interpolation and Extrapolation of Remote Sensing Images Based on Temporal Generation Model. Under review.
- Patent application CN202511188342.5: temporal-stability partitioning and Transformer-LaMA remote-sensing interpolation.
- Patent application 202511314521.9: masked-autoencoding interpolation for sea-surface reflectance products.
- Patent application CN202311554601.2: multi-vehicle urban sensing optimization.
- CEO, 深圳市视元引擎有限公司, Present.
- Research Assistant, Department of Electrical and Electronic Engineering, Southern University of Science and Technology, 2025-2026.
- M.S. in Computer Technology, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, 2025.
- B.S. in Computer Science and Technology, Shandong University of Science and Technology, 2019.
3D Gaussian Splatting Neural Rendering Novel-View Synthesis Monocular Depth PyTorch CUDA/GPU Systems Temporal Modeling Optimization
For research or industry collaboration inquiries, contact 522066928@qq.com.