|
Fusing RGB and Depth with Self-attention for Unknown Object Segmentation in Robotic Bin-picking |
|
Joosoon Lee(GIST, Korea), Seunghyeok Back, Taewon Kim, Sungho Shin, Sangjun Noh, Raeyoung Kang, Jongwon Kim, Kyoobin Lee(Gwangju Institute of Science and Technology, Korea) |
| We present a Synthetic RGB-D Fusion Mask R-CNN (SF Mask R-CNN) for unseen object instance segmentation. Our key idea is to fuse RGB and depth with a learnable spatial attention estimator, named Self-Attention-based Confidence map Estimator (SACE), in four scales upon a category-agnostic instance segmentation model. Our experiments showed the effectiveness of SACE in unseen object segmentation by achieving state-of-the-art performance. Also, we compared the feature maps varying the input modality and fusion method and showed that SACE could be helpful to learn a distinctive object features. |
|
|
Automatic Detection of Injection and Press Mold Parts on 2D Drawing Using Deep Neural Network |
|
Junseok Lee(GIST, Korea), Jongwon Kim, Jumi Park, Seunghyeok Back, Seongho Bak, Kyoobin Lee(Gwangju Institute of Science and Technology, Korea) |
| This paper proposes a deep learning–based method to automatically detect the injection mold parts (i.e., hook or boss) and press mold parts (i.e., DPS or Embo) in 3D CAD models of commercial TV. We first converted the 3D CAD models into 2D drawings and cropped them into a smaller image patch for the training efficiency of a deep neural network. Then, we found the position and type of mold parts using Cascade R-CNN and estimated the orientation of the detected mold parts using ResNet-50.Finally, we converted the 2D position of the mold parts to the 3D position of the original image. We expect our algorithms to contribute to faster industrial product design. |
|
|
Leveraging 3D Object Segmentation for Robust Visual Localization in Dynamic Environments |
|
Seongwon Lee(Yonsei University, Korea), HyungGi Jo(Jeonbuk National University, Korea) |
| Visual localization estimates the current 6-DOF pose using the monocular camera image. One of the main
factors that affects the performance of visual localization in the real environment is the existence of moving objects. In this paper, we propose a visual localization method less affected by moving objects. In order to reduce the influence of moving objects, the proposed method detects vehicles, two-wheeled vehicles, and pedestrians most commonly found in
object urban environments. Then, particle filter
dependent on segmented objects estimates 6-DOF pose. The weights of the particle filter are updated in the probabilistic framework using the segmentation results as additional measurements. |
|
|
Low-cost Selective Steering Structure Design for Counter-steering and In-place Rotation of Smart Farm Robots |
|
SooHwan Byeon, Jaebyung Park(Jeonbuk National University, Korea) |
| - A novel mobile robot steering structure is proposed for counter-steering and rotating in place.
- Car-like robots and differentially driven robots are easy to make, but they cause slip when they rotate.
- AWS (All-Wheel-Steer) robots with as many motors as the number of wheels can reduce slips and increase maneuverability, but their software and hardware systems are complex and expensive.
- Thus, we propose a more simple and inexpensive steering structure with high maneuverability like AWS.
- The proposed structure has only a single motor and a solenoid actuator to achieve high enough mobility to do tasks required in the smart farm. |
|
|
A Study of Paprika Disease Detection with YOLOv4 Model Using a Customed Pre-training Method |
|
HyungJun Jin(JeonBuk National University, Korea), HyongSuk Kim(, Intelligent Robots Research Center, Jeonbuk National University, Jeonju, South Korea, Korea) |
| In this study, we employ object detection technique to detect 5 paprika diseases in greenhouse, namely Blossom
end rot, Graymold, Powdery mildew, Spider mite, and Spotting disease. YOLOv4 model is used to detect the diseases in real-time with better detecting accuracy. Transfer learning is incorporated to enhance the detection performance where we employed 2 different methods to investigate the performance of the YOLOv4 backbone architecture. In first method, we only used the pre-trained YOLOv4 backbone on ImageNet classification dataset, whereas in second method, we tuned the pre-trained backbone weights (i.e. first method) with our own cropped paprika disease images. |
| |