Yoonki Cho

Visual Place Recognition (VPR) is a task of estimating the location of a query image, predominantly executed through image retrieval using learned global descriptors from a reference database of geo-tagged images. While recent approaches have aimed to improve the scalability of VPR training by leveraging classification loss as a proxy task, this objective discretizes the feature space into distinct class regions, often overlooking visual differences between classes. This discretization makes VPR systems particularly vulnerable to extreme visual changes such as lifelong variations. To remedy these problems, we propose a novel Class-Relational Objective (CRO) that transforms one-hot labels into soft labels by considering visual information of inter-class relations. We further enhance this method by dynamically adjusting the influence of CRO based on class weight magnitudes, which serve as an indicator of class stability as supported by derivative analysis. Comprehensive experiments across diverse benchmarks, ranging from general visual variations to extreme lifelong changes, demonstrate that our method outperforms existing methods. The source code will be made publicly available.

In this letter, we present real-time collision-free inverse kinematics (RCIK) that accurately performs consecutively provided six-degrees-of-freedom commands in environments containing static and dynamic obstacles. Our method is based on an optimization-based IK approach to generate IK candidates with high feasibility for the command. While checking various constraints (e.g., collision and joint velocity limits), we select the best configuration among generated IK candidates through our objective function, considering the continuity of joints and collision avoidance with obstacles. To avoid dynamic obstacles efficiently, we propose a novel, collision-cost prediction network (CCPN) that estimates collision costs using an occupancy grid updated from sensor data in real-time. We evaluate our method in three dynamic problems using a real robot, the Fetch manipulator, and four static problems using three different configurations of robots. We show that the proposed method successfully performs the consecutively given commands in real-time, mainly thanks to the collision-cost prediction network, while avoiding dynamic and static obstacles. The results of tested problems are available in the accompanying video.