Oh Won Jun

Proposed a novel Vision Transformer-based dual-branch model for depth completion, effectively fusing LiDAR and camera data. Achieved state-of-the-art results on NYUv2 and KITTI night scenes. (link)

Enhanced NICE-SLAM by incorporating uncertainty estimation into both photometric and geometric losses, resulting in improved robustness under challenging sensor conditions (link)

Proposed a framework that exclusively compresses xyz coordinates of 3DGS using a PCC model, enabling seamless integration with existing attribute compression methods. Fine-tuning the PCC decoder boosts PSNR and storage efficiency, as demonstrated on Synthetic NeRF datasets. (link)

Proposed an INR-based end-to-end image downscaling pipeline that adaptively predicts the optimal scale for each image to maximize detection performance under compression constraints. Achieved improved efficiency and accuracy over traditional and heuristic adaptive approaches on COCO. (link)

Built an OpenGL renderer displaying nine textured planets in a Cornell Box with real-time shadows. Used per-fragment Blinn-Phong shading and shadow ray-casting in GLSL for visually accurate shadow effects, including self-shadow avoidance. Implemented interactive camera controls and planet collision avoidance. (link)

Proposed a novel rendering method that visualizes the entire drawing history of a 3D object by dynamically controlling Gaussian primitive opacity using binary masks, built on top of existing models (Paints-Undo and Triplane-Gaussian) for drawing stage extraction and 3D reconstruction. (link)