Junsu Kim, 김준수
I am a MS student in 3D Vision & Robotics Lab, Department of Artificial Intelligence at UNIST, advised by Kyungdon Joo. During my MS, I worked as an intern at 42dot Inc. and at BIG (roBot Ingelligence Group) of Robot Institute at CMU, advised by Jean Oh. I received my BS from the Department of Automobile and IT Convergence at Kookmin University, advised by Sanghun Lee.
Research
🚀 My research interests lie in the fields of 3D computer vision and machine learning, particularly 3D vision for autonomous driving and robot arm, but not limited to.
VPOcc: Exploiting Vanishing Point for Monocular 3D Semantic Occupancy Prediction
Under review [ Paper ]
By utilizing a vanishing point in monocular 3D semantic occupancy prediction, this approach aims to enhance performance by incorporating camera perspective geometry into the network.
3D Tooth Loss Detection and Dataset from Intraoral Scans for Automation in Dental Implant Placement
ACCAS 2024
Estimating 3D bounding box of missing tooth loss area using point cloud of 3D dental scans.
Monocular Fisheye 3D Object Detection
On-going
We propose a monocular 3D object detection model for fisheye cameras using the geometric concept of the fisheye camera.
Project
SoftBiT: Soft Bimanual Teleoperation with Proprioceptive Visual Augmentation
Uksang Yoo, Tai Inui, Junsu Kim, Jeong Hun Lee, Peter Schaldenbrand, Kyungdon Joo, Nathaniel Dennler, Jonathan Francis, Jeffrey Ichnowski, Jean Oh
XRTC Symposium 2024
Building on a basic surround-view system of pinhole cameras, this approach incorporates fisheye cameras to enhance the performance of multi-view 3D object detection.
Enhancing Surround-view 3D Object Detection with Fisheye Camera
Junsu Kim
Building on a basic surround-view system of pinhole cameras, this approach incorporates fisheye cameras to enhance the performance of multi-view 3D object detection.
Exploring Point Cloud Attention: Unraveling its Impact on Point Cloud Shape Completion
Junsu Kim, Minje Kim, Dongjun Gu
Principles of Deep Learning Course Project, 2023
In a point cloud completion method using Transformers, we analyze attention tendency in the encoder and decoder, and based on this, propose a method using Minimum Density Sampling (MDS).
DeepPAVE: Deep learning-based Personalized Autonomous VehiclE
Team HAVI (Project leader)
The 9th Cloud ProgrammingWorld Cup, 2021 [ Poster ] [ Certification ]
Building a personalized autonomous driving system via socket communication between Virtual Simulation, Deep Learning Model, K7 Simulator, and GUI.
Deep-Learning-based Personalization of Driving Behavior of Autonomous Vehicles
Junsu Kim, Jeongsu Sun, Seungyoon Lee, Ahhyeon Lee, Juhee Lee, Sang Hun Lee
KSAE 2021 Annual autumn conference & exhibition [ Paper ] [ Poster ] [ Certification ] [ Link ]
Using simulations and ADAS, we developed a model for personalized autonomous driving by predicting lane change timing and trajectory from individual driving data.