MultiFlyFour sensing modalities. One shared semantic view of the world.
A real-world aerial dataset connecting RGB, thermal, LiDAR, and radar through geometry-driven semantic label transfer.
Research Engineer · Fraunhofer IVI
Taehyoung
Kim.
I’m a research engineer working on robotic perception, 3D scene understanding, and aerial autonomy. I build systems that connect what a robot sees with the world around it.
A closer look / 01CVPR 2025
Semantic Similarity Propagation
↗Seeing consistently.
Frame after frame.
KD-SSP crop from the paper’s video, playing inline. Open the full comparison to see all four methods. Source & context
Robotic perception3D scene understandingNear-ground aerial autonomy
01 / A little about me
I’m interested in how
robots understand the world.
Taehyoung Kim.
I’m a Research Engineer at Fraunhofer IVI, working on perception systems for autonomous aerial robots. My research spans video understanding, multimodal sensing, and semantic 3D mapping.
I’m interested in the space between a promising research idea and a system that works beyond the benchmark. This site brings together my publications, experiments, and, over time, notes on what I’m learning and building.
More about my background02 / What I’ve been working on
Recent activity.
Recent papers and project updates. A small record of what I’m working on and sharing.
Publication
MultiFly preprint
A real-world aerial dataset connecting RGB, thermal, LiDAR, and radar through geometry-driven semantic label transfer.
Publication
VoxelFix preprint
Graph-based correction of semantic labels in completed voxel maps, keeping geometry and occupancy fixed.
Publication
MissionBench preprint
Mission-level evaluation of multimodal language-model agents for aerial tasks, including qualitative failure analysis.
Publication
FlexPath preprint
Learned semantic path priors connect image understanding with image-based planning.
03 / Research
Selected research.
VoxelFixBetter semantic maps. Same geometry.
Graph-based correction of semantic labels in completed voxel maps, keeping geometry and occupancy fixed.
SSPSeeing consistently, frame after frame.
Semantic similarity propagation improves temporal consistency in video semantic segmentation for autonomous flight.
MissionBenchFrom individual actions to complete missions.
Mission-level evaluation of multimodal language-model agents for aerial tasks, including qualitative failure analysis.
FlexPathLearning where a path could go.
Learned semantic path priors connect image understanding with image-based planning.
Collaborative research, with full author credits on every paper.Full publication record
04 / The questions that connect it
Perceive. Represent.
Support decisions.
Three connected research directions. Distinct projects, shared questions about reliable robotic perception.
Real-time
perception
How can a robot understand what it sees, consistently over time and across sensing modalities?
Video perception / SSP3D scene
understanding
How can semantic geometry become a more reliable representation of the world?
Semantic mapping / VoxelFixAerial
autonomy
How can perception support planning, and how should we evaluate an agent’s complete mission?
Mission evaluation / MissionBench05 / Look beyond the figure
Same world.
A clearer map.
Explore what changes when semantic labels are corrected, while the underlying geometry stays fixed.
The VoxelFix viewer places input, corrected, and ground-truth maps side by side, with synchronized cameras and real scene data.
Explore the voxel maps Interactive 3D opens on the project page.
06 / Beyond the paper
Research meets
real systems.
My work sits at the intersection of computer vision, semantic mapping, and aerial robotics.
At Fraunhofer IVI, I work in a collaborative research environment where robotic perception connects with practical autonomous systems.
Applied research contextIdeas are better in conversation
Let’s talk about
what comes next.
Research, collaboration, or an interesting technical question.
I’d be happy to hear from you.