Research

Reliable AI for 3D visual perception

One question runs through my work: what can be recovered about a three-dimensional scene when the images available do not determine the answer? Dense, well-distributed capture is a laboratory condition. The settings that matter most, such as a camera on a vehicle, a drone following a flight path, or a handheld capture of a space walked through once, fix the coverage in advance.

Reconstruction from sparse views

Recovering geometry and appearance from a few, unevenly distributed views, using stereo priors, targeted densification and optimization that stays stable when the data are insufficient.

3D detection and scene understanding

Recognizing and locating objects inside reconstructed scenes, with volumetric learning that stays robust when views are sparse and noisy.

Efficient, anti-aliased rendering

Reducing model size and training cost without losing quality, and keeping rendering stable at scales and distances outside the fitted regime.

Trustworthy AI and uncertainty

Physics-informed learning with calibrated uncertainty for engineering prediction, and models that report where their own evidence is thin.

Research systems

Complete systems, each released publicly

Each system targets a different point of failure, and each is released with documentation and the configurations needed to reproduce every reported number.

APSIPA ASC 2026

ATW-GS

Anisotropic transmittance windows for anti-aliased 3D Gaussian splatting

Per-pixel transmittance is modelled as an oriented Gaussian window, so rendering stays free of aliasing when a scene is viewed at resolutions other than those used in training.

Anti-aliasingMulti-scale rendering3D Gaussian splattingDetails

2026

SMAGaussian

Stabilizing joint optimization with weight averaging and multi-resolution supervision for egocentric 3D Gaussian splatting

Fitting geometry and appearance jointly from egocentric video is unstable. SMAGaussian stabilizes the optimization itself, with weight averaging and multi-resolution supervision, adding no learnable parameters and no inference cost.

3D Gaussian splattingEgocentric videoOptimizationDetails

2026

FastGS-Adaptive

Structure-aware, view-consistent densification for accelerated 3D Gaussian splatting

Densification is directed toward primitives that carry structural error instead of those covering smooth surfaces, improving quality on all three standard benchmarks without changing the rasterizer.

EfficiencyDensification3D Gaussian splattingDetails

2026

StereoSplat

Enhanced sparse-view 3D Gaussian splatting via stereo-depth bootstrapping

Dense stereo depth becomes explicit scene geometry: depth maps are lifted into seed points and injected as new Gaussians, filtered by a left-right consistency check, while virtual views add photometric supervision.

Sparse viewsStereo depthSurface reconstructionDetails

Neurocomputing · 2026

NeRF-UAVeL

Unified attention-driven volumetric learning for robust NeRF-based 3D object detection

3D object detection from posed multi-view images through neural radiance fields. Four attention modules for volumetric feature learning keep detection accurate when views are sparse and noisy.

Neural radiance fields3D object detectionAttentionDetails

Next direction

AI that knows when it does not know

Uncertainty is usually computed during reconstruction and then discarded before recognition. My next line of work treats reconstruction, recognition and the confidence attached to both as one estimation problem, so that a system can report where its own evidence is thin instead of answering anyway.

This matters wherever a 3D model feeds a decision: robotics, digital twins, autonomous systems and infrastructure monitoring. It also divides naturally into projects of student size, from calibrating a single downstream decision to defining what calibration means for a prediction spread over space instead of a class label.

Collaborations

The Hong Kong Polytechnic University
Kennesaw State University
Idaho State University

PublicationsBrowse all publications