Dongguk University
← Back to Research
02

Spatial AI

We work on visual localization, 3D reconstruction, scanning, and scene understanding, enabling robots to build and continually update accurate 3D representations of their surroundings.

Spatial AI: real-time high-precision 3D mapping results

Visual SLAM

  • Localization: robustly extracts dense correspondences under extreme conditions — lighting change, repetitive patterns, textureless regions — for precise camera pose estimation.
  • Mapping: real-time, high-precision 3D reconstruction via online multi-view stereo.
2 top-tier AI conference papers (AAAI-23, CVPR-24) · 1 top journal, IEEE TIP (top 2% JCR)

Image-based real-time, high-precision camera localization and 3D mapping technology.

  • TopicFM: Robust and Interpretable Topic-Assisted Feature Matching — AAAI 2023
  • Learning to Produce Semi-Dense Correspondences for Visual Localization — CVPR 2024
  • TopicFM+: Boosting Accuracy and Efficiency of Topic-Assisted Feature Matching — IEEE TIP, 2024
Feature matching results in extreme environments

MVS & Neural Rendering

  • Multi-view stereo (MVS): high-precision 3D reconstruction using a depth foundation model, ranking 6th on the Tanks and Temples benchmark.
  • Online 3DGS reconstruction: generates and updates a 3D Gaussian map in real time for photorealistic, high-quality rendering.
  • Large-scale urban reconstruction: rapidly reconstructs wide urban areas in 3D, using a video diffusion model to synthesize novel-view images and fill in data-sparse regions.
3 top-tier conference papers (ICLR-22, IJCAI-25, CVPR-26) · 1 top journal, Pattern Recognition (top 8.8%) · 3 SCIE papers · 3 technology transfers

3D reconstruction technology based on multi-view stereo (MVS) and 3D Gaussian Splatting (3DGS).

  • Curvature-Guided Dynamic Scale Networks for Multi-View Stereo — ICLR 2022
  • Online 3D Gaussian Splatting Modeling with Novel View Selection — IJCAI 2025
  • PR-IQA: Partial-Reference Image Quality Assessment for Diffusion-Based Novel View Synthesis — CVPR 2026
  • Prior Depth-Based Multi-View Stereo Network for Online 3D Model Reconstruction — Pattern Recognition, 2023
MVS and Neural Rendering results

3D Scanning

Performs real-time 3D reconstruction from drone aerial imagery (online MVS), instantly identifies weakly-reconstructed regions, and plans an optimal re-scanning path — enabling complete 3D scanning of large-scale structures in a single flight.

4 top-tier robotics conference papers (ICRA-17/18/20, IROS-26) · 1 top journal, IEEE T-RO (top 7.3%) · 4 SCIE papers

Optimal 3D scanning path-planning technology.

  • Active 3D Modeling via Online Multi-View Stereo — ICRA 2020
  • Online Coverage and Inspection Planning for 3D Modeling — Autonomous Robots, 2020
  • View Path Planning via Online Multi-View Stereo for 3D Modeling of Large-Scale Structures — IEEE T-RO, 2022
  • Online Explore-Then-Exploit UAV Scanning with Budgeted Return Re-Scanning — IROS 2026
3D Scanning results

3D Scene Graphs

We propose a feed-forward framework that combines LLM-generated spatial constraints with best-view visual confidence to build verified 3D scene graphs consistent with the real environment — filtering false-positive objects without costly iterative updates, and generating reliable subtask sequences from verified objects only.

Spatial-constraint-based 3D scene graph and task-planning technology.

3D Scene Graphs results

Multi-Object Tracking

On-going research.

Multi-Object Tracking results
← Back to Research