Nov 2025 - Present: Talent Program Intern, Horizon Robotics

• Work on 3D reconstruction foundation models, with a focus on feed-forward multi-view geometry reasoning, cross-view communication, and high-resolution reconstruction.
• Developed ATSR, a block-level top-k sparse attention design that reduces redundant cross-view communication and improves Pose AUC@30 from 0.879 to 0.891 with 8.45x acceleration under 300-view evaluation.
• Developed GeoWeave, a selective cross-view communication mechanism for robust pose and point-cloud reconstruction under weak overlap and distracting views.
• Explored Hybrid-VGGT, a high-resolution architecture with a low-resolution global backbone and a high-resolution HDE branch for depth and point-cloud quality.
• Built mixed training data from 17 public datasets, unified evaluation scripts, business-data adaptation pipelines, and synthetic-data support for detailed object reconstruction.

3D Vision: Reconstruction and Scene Understanding

• Sparse-view 3D reconstruction: study initialization quality as a dominant factor for sparse-view 3DGS; proposed low-frequency-aware SfM, 3DGS self-initialization, and point-cloud consistency filtering.
• Object-level understanding in large 3D scenes: proposed Gaussian-segmentation guided object extraction and personalized diffusion priors for occluded or invisible regions.
• Monocular 3D human pose estimation: designed pose-guided attention and adaptive feature selection to leverage image cues for 2D-to-3D pose lifting.

Controllable Visual Generation and Resolution Extrapolation

• Analyzed implicit positional encoding in U-Net diffusion and proposed a training-free strategy for high-resolution image generation.
• Studied resolution scalability in Diffusion Transformers and FLUX-like models through attention receptive fields and frequency-aware detail preservation.
• Worked on image-only in-domain generation adaptation, controllable-generation evaluation, and a survey of text-to-image diffusion control.