Seattle, WA, United States.

University of Maryland, College Park, MD, United States.

• Proposed SLURPP, a single-stage latent diffusion model with dual-branch tailored for physically accurate scattering medium decomposition, achieving ~3dB PSNR gain and 200× speedup for underwater image restoration (accepted by TPAMI).
• Proposed a point-based normal flow estimator using a local point cloud encoder to predict per-event flow, enabling sharp, robust, and transferable predictions with uncertainty quantification, and supporting an IMU-based egomotion solver for challenging scenarios (accepted by ICCV 2025).
• Proposed ViewActive, a novel framework for active viewpoint optimization that mimics human-like spatial reasoning to enhance robotic perception (accepted by IROS 2025).
• Proposed Event3DGS, an event-based 3D Gaussian Splatting method that achieves state-of-the-art reconstruction quality and significantly accelerates training and rendering (accepted by CoRL 2024).
• Proposed MARVIS, a cutting-edge solution for real-virtual image segmentation near water surfaces, effectively leveraging synthetic data and domain-invariant features (accepted by IROS 2024).
• Teaching Assistant for CMSC426 Computer Vision.

Dolby Laboratories, Sunnyvale, CA, United States.

• Proposed a 3D live broadcasting solution for large-scale sports, integrating a 3D tracking foundation model prior to 4D sparse-view inverse rendering for high-fidelity scene reconstruction (Patent pending).

University of Florida · RoboPI Lab, Gainesville, FL, United States.

• Proposed AquaFuse, a physics-based waterbody fusion method for underwater imagery that preserves depth and object geometry, enabling geometry-consistent data augmentation and accurate 3D view synthesis with 94% depth and 90–95% structural similarity (accepted by RA-L).
• Proposed a 3D underwater reconstruction pipeline using pixel-wise depth guidance for sparse-to-semi-dense point clouds, reducing 2D feature reliance and achieving 3× faster inference with higher reconstruction accuracy (Best Paper Award at IEEE CAI 2023).
• Formulated a robust and efficient monocular depth estimation model named UDepth, by incorporating underwater domain knowledge into its supervised learning pipeline (accepted by ICRA 2023).

Remote Sensing Laboratory, University of Florida, Gainesville, FL, United States.

• Generate complex 3D models programmatically.
• Connect the program with the database to realize the dynamic real-time 3D model.
• 3 papers accepted by IGARSS 2023, IEEE TGRS and IEEE JSTARS.

Vobile, Santa Clara, CA, United States.

• Developed and implemented a learning-based video retrieval system based on global feature and local feature fusion. And also wrote the user manual and targeted model performance optimization guidelines document of the system.
• Conducted a number of qualitative phase-shift auditory tests and found a relationship between the phase-shift cases and the psychoacoustic model. Upgraded the audio fingerprint encoding algorithm based on the classic psychoacoustic model. The upgraded algorithm can encode not only the sound pressure level of the audio fingerprint but also the threshold of its phase shift (implemented in C and MATLAB).