2001.04.02
<aside> 🌱 I have a strong interest and passion for AI technology development and research, particularly in the areas of 3D vision and human-centered AI.
</aside>
(M.S./PhD integrated) Seoul National University (SNU)
(B.S.) Gwangju Institute of Science and Technology (GIST)
Research interest
Contact
Github: https://github.com/SojeongKim-42
Now in an intergated course(M.S.) at SNU AIoT Lab. Developed fundamentals of 3D vision, assisting in some research on building a 3D dataset. I am particularly interested in leveraging egocentric perspectives to enhance 3D vision systems. Currently exploring pressure feature learning based on vision hand-object interaction image.
Interned at GIST Computer Graphics Lab and conducted bachelor’s thesis research ****at the intersection of neural rendering and image processing. Applied a cross-bilateral filter to quickly rendered low-spp images guided by NeRF outputs, achieving significant RSE reductions. This approach effectively improved both rendering time and image quality, by just inferencing from a pretrained NeRF model.
Created data classification rules using regular expressions and SQL queries. Identified inefficiencies in the internal management site and took initiative to fix bugs and add new features using React and Scala. Proposed the adoption of ML to detect misclassifications and led the cross-team collaboration to integrate it.
Introduced EgoTrack, a robust 6D object tracking framework for egocentric videos under severe visual degradation. Instead of relying on a single frame-wise estimate, EgoTrack preserves multiple pose hypotheses and dynamically reasons over them across time to recover accurate and consistent object trajectories. I contributed to research design and performance improvement strategies, experiments, and paper writing.
Introduced EgoXpose, a scene-aware egocentric 3D human motion reconstruction framework that learns scene context from exocentric supervision without requiring explicit 3D scene geometry at inference time. EgoXpose improves the physical consistency of reconstructed motion and introduces a 3D scene penetration metric for evaluating body-scene interactions. I contributed to research design, visualization, EgoEgo baseline experiments, and the 3D scene reconstruction and penetration evaluation pipeline.