previous day
next day
all days

View: session overviewtalk overview

13:00-18:00 Session 14: Robot Seminar
Shape Restoration by Shadow Information and Photometric Stereo

ABSTRACT. Virtual Reality (VR) systems are become popular in recent years. Along with that, studies that capture 3D objects from the real world have been studied. 3D objects however have large and complex data. In this paper, we propose a novel method that use the shadow information and Photometric stereo with their surface from one point of view. The experimental results show that the proposed method performs well in accuracy.

Characteristics based visual servo for 6DOF robot arm control

ABSTRACT. Visual servo is a method for robot arm motion control. It is controlled by the end effector velocity that is the result in calculating the internal Jacobi matrix and vector of feature error. In general, automatic robotic task require high quality sensor that can measure a 3-dimential distance, and do calibration in order to suit the sensor frame and robot frame in Euclidean space. In this paper, we only use RGB camera and not requiring the calibration in sensor frame, so it is simpler than any other automatic motion methods. The proposed characteristics based visual servo method has varying the hyper parameter, and show the effectiveness for indicating the precision of pose error both simulation and actual environments.

Underwater Image Super-resolution Using SRCNN

ABSTRACT. In recent years, energy minerals have become more important due to the rapid industrialization worldwide. Due to the rapid industrialization on a global scale, there is a shortage of mineral resources, and there are more opportunities to rely on alternative energy sources. Therefore, the exploration of marine resources, which are abundant in the ocean, is being promoted. However, it is dangerous and impractical for humans to actually dive and search for marine resources by hand. Therefore, it is possible to proceed with underwater exploration safely by having a robot do the work instead. Robots have been used as a mainstream search tool in the underwater environment due to the existence of various hazardous environmental conditions. However, there are several problems associated with robot control in underwater environments, one of which is poor visibility in the water. One of the problems is the poor visibility in the water. To improve the visibility in the water, we are trying to increase the resolution of underwater images by using super-resolution technology. In this paper, we conduct experiments using SRCNN, which is a basic super-resolution technique for underwater images. In addition, we investigate the effectiveness of "Mish", which has been attracting attention in recent years for its potential to surpass the performance of "ReLU", although "ReLU" is a typical activation function of neural networks, on SRCNN.

Single Image Reflection Removal Using Meta-learning

ABSTRACT. In recent years, reflection is a kind of noise in images which is frequently generated by reflections from windows, glasses and so on when you take pictures or movies. The reflection do not only degrade the image quality, but also affects computer vision tasks such as object detection and segmentation. In SIRR, learning models are often used because various patterns of reflection are possible, and the versatility of the model is required. In this study, we propose a deep learning model for SIRR. There are two problems with the conventional SIRR using deep learning models. The assumed scenes of reflection are vary, and there is little training data because it is difficult to obtain true values. In this study, we focus on the latter and propose an SIRR based on meta-learning . In this study, we adopt MAML, which is one of the methods of meta-learning. In this study, we propose an SIRR using a deep learning model with MAML, which is one of the methods of meta-learning. The deep learning model includes the Iterative Boost Convolutional LSTM Network (IBCLN) is adopted as the deep learning model. Proposed method improve accuracy compared with conventional method of state-of-the-art result in SIRR.

Image Quality Improvement Method Using Local Adaptive Neighborhood-based Dark Channel Prior

ABSTRACT. In-vehicle cameras and surveillance cameras are used in many situations in our daily lives. Visibility degradation in foggy environments is caused by the scattering of reflected light from real objects by minute water droplets or fog in the medium through which light passes. The degree of degradation depends on the density of suspended microparticles existing between the observed object and the observation point in the medium. In general, the farther the object is from the camera, the more it is affected by the fog. The purpose of image de-fogging is to improve the clarity of an object by removing the effects of fog in the image.

PointNet++: Deep Hierarchical Feature Learning
NDT-Transformer: Large-Scale 3D Point Cloud Localization
Generative Robotic Grasping
Learning Synergies between Pushing and Grasping
Go-ICP: Solving 3D Registration
Modelless Grasping Points Estimation
Sim-to-Real Reinforcement Learning
PointNet: Deep Learning on Point Sets