Perception
Segmentation, grounding, and reasoning in 3D scenes.
News
- Jun 2026 One paper has been accepted to ECCV 2026.
- Mar 2026 One paper has been accepted to CVPR 2026.
Research
My research starts from 3D scene understanding including perception, reconstruction and rendering. Nowadays, I am working on 3D scene interaction where embodied agents do reasoning and exploration.
Reconstruction and Rendering
Pointcloud, voxel, and Gaussian for representing 3D scenes
Embodied Agents
Action and exploration in real-world or 3D simulators.
Selected Publications
Representative publications are shown first, followed by earlier work in 3D vision.
SplatReasoner: Enhancing Embodied Reasoning and Grounding by Novel View Synthesis
Novel-view synthesis improves the reasoning capabilities of VLMs in 3D scenes.
OpenVoxel: Training-Free Grouping and Captioning Voxels for Open-Vocabulary 3D Scene Understanding
Any VLMs or LLMs can now reason over 3D scenes.
DVSM: Decoder-only View Synthesis Model Done Right
3D scene is encoded as KV-cache by our decoder-only model.
MV-SAM: Multi-view Promptable Segmentation using Pointmap Guidance
View-consistent segmentation using pointmaps
Quantile Rendering: Efficiently Embedding High-dimensional Feature on 3D Gaussian Splatting
Efficient color/feature rendering algorithm from 3D Gaussians.
SVRecon: Sparse Voxel Rasterization for Surface Reconstruction
Voxel-wise associations among parent-child and sibling voxel groups produce smoother surface reconstruction.
Dr. Splat: Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration
Language knowledge is distilled into 3D Gaussian Splatting for direct referring.
Mosaic3D: Foundation Dataset and Model for Open-Vocabulary 3D Segmentation
Data generation pipline from 2D images to 3D mask-text.
Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field Rendering
Rasterization is extended into sparse voxel representations for real-time image rendering.
Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction
Spacetime surface regularization for 4D surface reconstruction and dynamic scene rendering.
PointMixer: MLP-Mixer for Point Cloud Understanding
A new MLP-only architecture for 3D point cloud understanding.
Deep Point Cloud Reconstruction
Point cloud upsampling and denoising are performed jointly with strong generalization.
VolumeFusion: Deep Depth Fusion for 3D Scene Reconstruction
A deep learning based depth fusion algorithm for indoor scene reconstruction.
Volumetric Propagation Network: Stereo-LiDAR Fusion for Long-Range Depth Estimation
Dense depth estimation by fusing LiDAR points and stereo images.
Stereo Object Matching Network
Stereo matching is guided by object-level 3D context inside the cost volume.
Segment2Regress: Monocular 3D Vehicle Localization in Two Stages
Monocular 3D vehicle localization is decomposed into road-plane segment prediction and 3D regression.
Mentorship and Collaboration
Research Intern Mentorship
Usually, I have a chance to hire one intern at each year. Feel free to send me an email if you are interested in this position.
- 2026Kim Yu-Ji, POSTECH -> with me
- 2024Yoonwoo Jeong, POSTECH -> with me -> Meta internship
- 2023Chunghyun Park, POSTECH -> with me
Academic Collaboration
University collaborations are always welcome. I am happy to join academic collaborations when a student’s research goals matches with my interests. Authorship order is not my main concern; I value the opportunity to contribute to meaningful and high-quality projects.
- 2026Open (feel free to reach out if interested!)
- 2025Seunghun Oh, Seoul National University -> with me -> UCLA
- 2025Kim Yu-Ji, POSTECH -> with me -> NVIDIA internship
- 2024Kim Jun-Seong, POSTECH -> with me -> Huawei/Meta internship