I am an MSc student in Artificial Intelligence and Robotics at The Chinese University of Hong Kong, Shenzhen, advised by Prof. Rui Huang. I work on 3D perception that robots can rely on. My thesis studies uncertainty calibration for self-supervised monocular depth, and my capstone made Gaussian-based 3D occupancy differentiable and 7.2× faster.

Going forward, I want to build 3D scene representations that stay fast under limited compute, and perception that knows when it should not be trusted. I am applying to PhD programs for Fall 2027.

Homepage · CV · ijunsuk9797@gmail.com

Range-aware uncertainty calibration in self-supervised monocular depth (MSc thesis, Jan 2026 – present)
Manuscript in preparation.

LocalAggTR: Differentiable voxelization for weakly supervised 3D occupancy (Sep 2025 – Dec 2025)
Replaced GaussTR's heuristic, non-differentiable voxelizer with GaussianFormer's Local Aggregation CUDA operator, making the Gaussian-to-voxel step differentiable.
→ 7.2× faster inference at comparable mIoU on nuScenes val, with the same checkpoint and head. Adopted by the lab's 3D Gaussian Splatting project for semi-supervised fine-tuning.

Geometric vs. Generative 3D Augmentation: Low-data point cloud classification (Jan 2025 – May 2025)
Tested whether Point-E text-to-3D augmentation helps PointNet on ModelNet40 compared with geometric augmentation.
→ Geometric augmentation raised accuracy by 6.5 points, while text-to-3D generative augmentation did not help.

Point Cloud Density and 3D Object Detection (Jun 2025 – Jul 2025)
Measured how a pretrained PointPillars degrades under reduced LiDAR density on nuScenes-mini, and whether simple upsampling recovers it.
→ Voxel-grid thinning to one fifth of the points cost 2.1% of mAP, while random thinning to half cost about 4× more.