Research

Research interest in computer graphics, perceptual rendering, computer vision and machine learning. Current primary focus in neural rendering, inverse rendering, multimodal reconstruction and novel view synthesis. Representative papers are highlighted.

Extracting features of feed-forward geometry transformers (like VGGT) into a consolidated grid-structured latent space and directly decoding a 3D representation.

A performance survey, analysis, integration, and extension of Gaussian Splatting training optimizations, leading up to a 5× speedup in training time without loss of quality.

Gaussian Splatting for multi-spectral images (e.g. red-edge, near-infrared) using a joint color-band feature representation and a neural emission function.

Large-scale radiance field rendering with neural point octrees for hierarchical point rendering, enabling efficient level of detail selection and real-time rendering of complex scenes.

Completion of LiDAR point clouds with photometric Gaussian Splatting reconstruction, based on region ambiguity.

Exploration of the integration of multi-spectral capturings (e.g. near-infrared) in Gaussian Splatting systems through investigating training regimes.        (*) Denotes equal contribution.

Dense and multi-view consistent depth maps optimized from a monocular depth estimation initialization.

Faster implicit neural point cloud (INPC) training and rendering with less VRAM and better quality.

Foveated radiance field rendering with smooth Gaussian peripheral renderings and neural point splatting for crisp foveal rendering.

A perspective-correct and view-consistent approach for 3D Gaussian splatting, accelerated through hybrid transparency.

Implicit volume-based scene reconstruction combined with point rendering for high detail novel view synthesis.

Rendering and optimizing neural point clouds via trilinear point splatting and tiny neural networks for fast and accurate novel view synthesis.

Improving point-based novel view synthesis quality by completing point cloud with 3D error volumes from 2D error maps.

Point-based novel view synthesis without per scene preprocessing or training using neural image based rendering.        (*) Denotes equal contribution.

Novel view synthesis on live RGB-D streams with feedback during capturing using on-the-fly surface integration, robust to slam loop closures.

Fast parallax-based light field display rendering through device-tailored projection matrices, reducing fragment overhead.

Point-based differentiable one-pixel point rendering for automatic scene calibration and high-quality novel view synthesis.

Foveated Rendering via temporal forward reprojection for faster VR rendering, exploiting the visual acuity falloff.

A multi-layered depth of field method for simulating partial occlusion in real-time rendering.

Exploration of generative programming for efficient CUDA kernels in depth of field rendering.

A tiled splatting approach for fast depth of field rendering.