-
Kazuyuki ASAI, Shigeru AKAMATSU
Session ID: 21-03-01
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
We investigated whether there is any causal relationship between attractiveness of face images perceived by humans and dynamic characteristics of their eye movement during the judgement. Hidden Markov model (HMM), a machine learning method suitable for machine learning for recognition of time-series data, was applied to analyze dynamic properties of gaze motion, and characteristics of spatial distribution and stationary period of gaze points were analyzed by means of heat map representation. As a result, it was found that the points of gaze have a tendency to stay within smaller areas for a longer period for the subjects who evaluated the target faces as highly attractive. On the other hand, in case when the subjects evaluated the faces as un-attractive, the points of gaze had a tendency to move around in shorter periods. Such experimental results may indicate that gaze motion pattern generated while making attractive judgements has more inherent and discriminative characteristics compared to the case of making un-attractive judgements.
View full abstract
-
Yuki Ono, Kazuhisa Yanaka
Session ID: 21-03-02
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
The “Rotating Snakes” illusion is an illusion presented by Kitaoka, and although it is a still image, it appears to move. However, the mechanism has not been fully elucidated yet. We conducted an experiment in which the original image and the gradation inversion image were presented repeatedly for about 100 ms each, and found a phenomenon in which the amount of illusion increased. At the same time, in the two repeated presentations, the direction of rotation may appear to be unidirectional or changed, but it was confirmed that there is a strong tendency for the direction of rotation to appear to rotate in the same direction as the normal rotating snakes illusion.
View full abstract
-
Ryohei YAMAZAK, Shigeru AKAMATSU
Session ID: 21-03-03
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
The purpose of this study is to verify the usefulness of using features that combine skeletal information and distance information for motion identification in video images. I measured the skeletal coordinates and the distance information around the palms, ankles, and head during the movements of radio gymnastics, and calculated the direction vectors of the arms and legs and the angles of the elbows, knees, and hips from the skeletal information. We used the hidden Markov model for identification. As a result, accuracy of the combined feature of skeletal information and distance information increased by 5.23% compared with that of the skeletal information alone and 16.19% compared with that of the distance information alone.
View full abstract
-
Yushi Okada, Chanjin Seo, Motofumi Taniguch, Kazuyuki Kanosue, Hiroyuk ...
Session ID: 21-03-04
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Skiing is a difficult sport for beginners to start compared to other sports because of the severe restriction of time and place, and the high risk of injuries such as falling. Therefore, a ski training system that integrates VR and hardware technology is developed. In this paper, we evaluate its effectiveness through experiments. Specifically, we verify and compare the training effects of two systems with different feedback methods: “a system that feeds back the changes in somatosensory perception as VR images” and “a system that feeds back the changes in somatosensory perception by increasing or decreasing the gauge”. By comparing these two systems, we examine whether the learning effect of the system can be improved by incorporating VR technology into the ski learning system, and obtained promising experimental results.
View full abstract
-
Daiki YAMAUCHI, Tsubasa TOMITAKA, Makoto J. HIRAYAMA
Session ID: 21-03-05
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Currently, the number of tourists is decreasing due to the spread of the new coronavirus infection. Hirakata City is no exception to this trend. As the coronavirus has reduced the number of people going out as well as going to other prefectures, there is a problem that people have fewer opportunities to come in contact with historical buildings and cultural assets. If we can create 3D models of historical buildings and cultural assets, we can place the 3D models on smartphones, tablets, and VR space, and we can touch historical buildings and cultural assets without going to the site. However, a 3D model of cultural properties in Hirakata City has not been created yet. However, there is no 3D model of cultural properties in Hirakata City. Therefore, we propose photogrammetry by drone as a method to create 3D models. In the proposed method, a drone is flown around an object to be measured and photographs are taken from various angles. After that, photogrammetry is performed by 3DF Zephyr, which is a software for photographic measurement, and a 3D model is generated. This method can perform photogrammetry on a wider range of objects than PIX4, a photo measurement software for drone mapping.
View full abstract
-
Yohei SAITO, Munetoshi IWAKIRI, HERNAN Aguirre, Kiyoshi TANAKA
Session ID: 21-03-06
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
The problem of estimating the transformation that makes the 3D point cloud data obtained from multiple viewpoints into an appropriate placement is called registration. When there are few overlap areas between two point clouds, the registration problem is difficult. Thus, we proposed a registration method that extracts a set of keypoint patches from a point cloud and estimates the overlap area using a genetic algorithm. However, the genetic operator of the proposed method is simple and inefficient, especially for a small set of overlapping points. Therefore, in this study, we introduce a new crossover method using the distance of keypoint patches.
View full abstract
-
Takashi Iwase, Munetoshi IWAKIRI, Kiyoshi TANAKA
Session ID: 21-03-07
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
A 3D point cloud is a representation of a 3D space with many discrete coordinate information by laser or photogrammetry. By processing 3D point clouds, features in 3D space can be directly analyzed, and various applications are expected. Although conventional contour extraction for 3D models is effective for modeled objects, it is not applicable to natural objects contained in remains. In this study, we propose a new algorithm for contour extraction based on the calculation of feature values using three kinds of indices (DoN, curvature, and ND-PCA) for 3D point cloud data obtained from actual remains, and verify its effectiveness through experiments.
View full abstract
-
Shinnosuke YAMAZAKI, Go IRIE, Ryosuke FURUTA, Yota YAMAMOTO, Yukinobu ...
Session ID: 21-03-08
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
A point cloud is one of the data representations of three dimensional object shapes. A point cloud represents the object shapes by surface points with their coordinates. The process of finding transformation of point clouds to align them into the common coordinate system is called registration. Dealing with the Partial-to-Partial (P2P) registration, where point clouds X and Y satisfy X ⊈ Y and Y ⊈ X , we propose a method for generating a mask to extract the intersection points X ∩ Y from two input point clouds X and Y. Experimental results using ModelNet40 show that the proposed method generates a mask that extracts the intersection points with higher accuracy than the conventional method.
View full abstract
-
Makoto ODAMAKI
Session ID: 21-03-09
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
THETA is the world’s first consumer 360-degree camera. This is widely used for business purposes such as virtual tours in real estate and construction markets. The omnidirectional image can expand computer vision techniques such as object detection, segmentation, localization, and 3d reconstruction and lead digital twin and spatial computing.
View full abstract
-
－Visualization of Regions of Interest Concerning Thrombus Formation by VGG16 and Grad-CAM as Well as Estimation of Thrombus Formation Factors by Light-GBM －
Sohta SAMPEI, Tadashi YAMAMOTO, Kiyotaka IWASAKI, Hiroshi NAGAHASHI, J ...
Session ID: 21-03-10
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Thrombus formation, in which blood coagulates in the left atrial appendage due to blood flow in the left atrium or atrial fibrillation, is one of the causes of cerebral infarction. Since surgical removal of the left atrial appendage is a heavy burden on the patient, if thrombus formation in the left atrial appendage can be predicted, prophylactic medication may prevent thrombus formation. However, the cause of the thrombus in the left atrial appendage has not been clarified, and the prediction method has not been established. In this study, we focus on blood flow information, left atrium shape, and four pulmonary veins as influential factors for thrombus formation, and explore and clarify the relationship between these factors and thrombus formation using the machine learning networks VGG16, Light-GBM, and Grad-CAM, which visualizes causes of the judgement done by VGG16.
View full abstract
-
Biao LI, Hiroki TAKAHASHI
Session ID: 21-03-11
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In Chinese medicine, tongue diagnosis is a simple and effective method to aid diagnosis by observing changes in the color and shape of the tongue. However, tongue diagnosis is costly to learn in terms of time and relies on the experienced judgement of the practitioner. The aim of this thesis is to discuss the correlation between tongue features and pathology through tongue image analysis. The images were first color-corrected and then the coating regions were segmented using the MLP (Muti-Layer Perceptron) algorithm. Based on the color features of the coating, the tongue image color attributes were evaluated and correlations with diabetes factors were discussed.
View full abstract
-
Yuya HONDA, Hiroki HATAKASHI
Session ID: 21-03-12
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In this paper, we propose a method to support early detection of hand-foot syndrome by identifying the manifestation of the syndrome using the fingertip region where symptoms appear in images of the affected area taken at home. First, we detect the color chart in the image and correct the color difference between each image by color correction. Next, we estimate the joint
points of the hand and extract the fingertip region of the target based on the joint points. Then, skin features that are indicators of symptom expression are extracted from the fingertip region. Finally, the extracted features are used to determine the onset of symptoms. As a result, the best performance of 0.82 for Recall and 0.66 for F-measure was obtained when SVM was used to classify the R channel using the feature size of the R channel relative to the B channel after color correction.
View full abstract
-
Kohei Takata, Yukihiro Michiwaki, Hiroshi Moriya, Takashi Ijiri
Session ID: 21-03-13
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In this study, we propose a semi-automatically segmentation method for bronchial regions from 4D computed tomography (4DCT) images. In our method, the user places control points with a tree structure similar to a bronchus. We then calculate paths between all pairs of parent and child control points and extract the bronchus shape by region growing using the paths as a seed. In this study, we search the paths by the shortest path algorithm considering the bronchial features. We also propose a region growing in which the user can control the growth size based on the depth of the tree structure. To demonstrate the feasibility of our method, we show multiple examples of bronchial region segmentation by using our prototype system.
View full abstract
-
Mei KODAMA
Session ID: 21-03-14
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In recent years, online video viewing has increased rapidly, and prevention of VIMS (visually induced motion sickness) is one of the important issues. So far, we have studied screen shake detection methods using motion histograms for screen shake. However, since these methods captures motion vector fluctuations due to screen shake as time changes in the motion histogram, there is a problem that if the threshold value is set small, the motion of the background or subject contained in the screen is erroneously detected and the detection accuracy decreases. Then, in this paper, the author proposes a new screen shake detection method by using a motion analysis that considers the frequency for the amount of motion of the entire screen in the motion histogram space. Specifically, he focuses on the amount of motion and the share of the highest frequency, and considers the change of motion distribution with these parameters greater than the threshold value. Next, this method detects whether the movement is the movement of the subject or the movement caused by the screen shake. In this proposed method, the improvement of screen shake detection accuracy is evaluated by simulation experiments and clarify the effectiveness of the proposed method.
View full abstract
-
Ryou MUTOU, Kazuya UEKI, Yuma SUZUKI, Takayuki HORI, Hideaki OKAMOTO
Session ID: 21-03-15
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In this paper, we propose a method for retrieving people wearing clothes of a specific category and color from a large-scale video database. In the proposed method, video retrieval is performed using the retrieval scores obtained from the module using image-language embedding method, which calculates the similarity between text and image features, and the module using instance segmentation and pose estimation method, which detects the region of a person and clothing. In order to verify the effectiveness of the proposed method, we conducted a retrieval experiment using a large-scale video database, and confirmed that the target video was in the top of the retrival results. When the weights of the scores obtained from the modules were changed, improvement of the average precision was confirmed.
View full abstract
-
Itsuki TANAKA, Munetoshi IWAKIRI, Kiyoshi TANAKA
Session ID: 21-03-16
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In this paper, we focus on a technique for reconstructing 3D point cloud data from video sequences, and investigate a method for selecting appropriate frames. The overlap rate between frames is calculated using optical flow as a feature value, and the suitable frame interval for reconstruction is determined. In addition, we compare several methods of frame selection considering the number of optical flows and frame information. The obtained experiments show that the frame selection method is effective in terms of the number of frames selected and the accuracy.
View full abstract
-
Hiroto KOBAYASHI, Go IRIE, Ryosuke FURUTA, Yota YAMAMOTO, Yukinobu TA ...
Session ID: 21-03-17
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
When watching a movie trailer or a sports highlight, you may want to jump to the full-length scene corresponding to the scene you are currently watching. It is time-consuming for a human to manually search for the scene you want to find by using skip and back, which impairs the video viewing experience and convenience. In this study, we propose a method for efficient retrieval of similar images from videos using deep reinforcement learning. We treat a typical scene as a query image and perform reinforcement learning of frame skip action. We improved the accuracy by devising the feature extraction module that takes into account the distance between the query image and the current frame’s image feature vector (feature distance), and by designing reward. As a result, the retrieval success rate is 83.0% and the number of steps is reduced by 48.2% compared to the fixed skip-width search with the same retrieval success rate in Multiview Pouring Dataset, a dataset of liquid pouring videos.
View full abstract
-
Tomoya Wakasugi, Raytchev Bisser, Kazufumi Kaneda, Masashi Baba
Session ID: 21-03-18
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
The color of objects captured in the sea is different from that of the observation on the ground due to the influences of scattering and absorption of light. Restoring the color of underwater images is important for investigating the material and surface condition of an object. In the previous report, we presented the image generation and color correction of the underwater image considering only the in-scattering component. In this report, we generate underwater images with various light sources taking into account all the scattering and absorption phenomena of light, and verify the effects of both light scattering and attenuation. The accuracy of the color correction results of the RGB images are also investigated. In addition, we propose a new color restoration method of RGB images taking into account spectral scattering and absorption coefficients, and verify the accuracy of the image restored by the method.
View full abstract
-
Kaito IKEHATA, Makoto FUJISAWA, Masahiko MIKAWA
Session ID: 21-03-19
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In this paper, we propose a simulation method for the formation of rocky coasts that are eroded by waves and sandy beaches that are formed by the transport of sediments such as sand. For rocky shores, SCAPE, which is a model for numerically calculating wave erosion and receding, is extended to three-dimensional terrain, and wave-cut notches in which the shore is scooped out by waves are also considered. For sandy beaches, a model is applied in which the suspended load layers and the bed load layers are considered separately to consider the effects of sand sedimentation. By combining these simulations, we conducted simulation experiments to reproduce wave erosion and sand transport/sedimentation and confirmed the effectiveness of the proposed method.
View full abstract
-
Ryosuke YAMASAKI, Makoto FUJISAWA, Masahiko MIKAWA
Session ID: 21-03-20
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In this paper, we propose a method to calculate the non-linear viscosity change accompanied by the gelatinization phenomenon of wheat flour suspension using fluid simulation. In the proposed method, a simulation is performed with considering the change in viscosity due to the gelatinization phenomenon that occurs when the suspension is mixed or heat is applied. The degree of gelatinization phenomenon is calculated from the change in thermal energy, assuming that the change in viscosity due to the gelatinization phenomenon is a state change with latent heat. Finally, we integrated the viscosity change due to the mixing and gelatinization phenomenon and conducted a simulation experiment assuming crepe making, and confirmed the effectiveness of the method.
View full abstract
-
Yucheng Qiu, Takuhiro Nishida, Takashi Ijiri
Session ID: 21-03-21
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Many studies have been presented to reconstruct insect specimens’ three-dimensional shapes and texture maps. However, their surface reflectance have rarely been reconstructed. In this study, we propose a method for reconstructing 3D shapes and reflectance characteristics of insect specimens. We take multiviewpoints focus-bracket images of a sample under two linear lights and reconstruct the 3D model of the sample by using a multi-view stereo method. From the 3D model and a series of photographs, we obtain a one-dimensional reflectance response function at each point on the 3D model, and estimate diffuse color, specular color, and surface roughness at the point. To illustrate the feasibility of our method, we performed digitization of multiple insect specimens with specular reflectance on their surfaces.
View full abstract
-
－Invisibility and Readability in Case of Checker Pattern－
Hiroshi UNNO, Kazutake UEHIRA
Session ID: 21-03-22
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
This paper describes a lighting technique which enables us to invisibly attach information onto a real object visible to the human eye by using illumination light and to extract the invisibly attached information from a video image of the object captured with a video camera. The information was embedded in the light beam by using temporal and spatial brightnessmodulation. Checker pattern was used as the information in this research. The pattern consists of squares having the same side length. The side length was changed to examine the invisibility and readability in terms of the basic unit of information. First, we assessed invisibility to the human eye of the information attached onto a real object. We also examined readability to a video camera of the attached information.
View full abstract
-
Akihisa Watanabe
Session ID: 21-03-23
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
The goal of learning disentangled representation is to obtain interpretable feature representations from data such as images. Although many disentanglement methods based on deep generative models such as variational autoencoders and adversarial generative models have been proposed, the experiments in this paper experimentally show the vulnerability of learning entangled representations to noise in images. In this paper, we propose a variational autoencoder-based disentanglement method that is robust against noise and self-supervised by latent representations for reconstructed images and verify its effectiveness through numerical experiments.
View full abstract
-
Kengo SATO, Kitahiro KANEDA, Keiichi IWAMURA
Session ID: 21-03-24
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
The distribution of easily imitated counterfeit products, such as food packaging, brand name tags, and pharmaceutical labels, has become a serious economic and safety concern. To address this issue, we propose a system for authenticity judgment of genuine and counterfeit products with high speed and accuracy, focusing on the physically unclonable function of an inkjet-printed code and a locally likely arrangement hashing (LLAH) system that performs high-speed image retrieval. In this study for the practical application of the system, we prepare a data set that assumes environmental changes that may occur in the real world when the system operates, and verify the robustness of the authenticity judgment system.
View full abstract
-
Ami HOSAKA, Kitahiro KANEDA, Keiichi IWAMURA
Session ID: 21-03-25
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
The distribution of easily imitated counterfeit products, such as food packaging, brand name tags, and pharmaceutical labels, has become a serious economic and safety concern. To address this issue, we propose a system for authenticity judgment of genuine and counterfeit products with high speed and accuracy, focusing on the physically unclonable function of an inkjet-printed code and a locally likely arrangement hashing (LLAH) system that performs high-speed image retrieval. In previous research, system for authenticity judgement has been build using inkjet printer for home use. To improve practicability, we used the dataset printed by an industrial inkjet printer used in the real world and a paper medium used as name tags and pharmaceutical labels. In this study for the practical application of the system, we verified that the proposed system has high discriminability and stability, based on highly accurate results obtained from the dataset assuming industrial use in the real world.
View full abstract
-
－Combination of Visual SLAM and Winch Odometry－
Taiki SUZUKI, Junji YAMATO, Hiroyuki ISHII, Jun OHYA, Atsuo TAKANISHI
Session ID: 21-03-26
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Vertical hole elevation robot has been developed to explore a lava cave at Marius Hills on the moon. This robot is suspended from the lunar surface by a tether and enters the cave to investigate the environment by climbing up and down the cave wall. In this process, the robot needs to estimate its own position. In this paper, we propose a localization method for the robot using a combination of Visual SLAM and winch odometry. The accuracy of the proposed method was verified by experiments using a simulation environment. The accuracy of the proposed method and the validity of the model used in the simulation were verified in experiments conducted in real environment. As a result of the experiments, we got a prospect that the proposed method can perform localization in the lunar vertical hole.
View full abstract
-
－calculation and usage of deformation process of 3D model－
Yuga Imaizumi
Session ID: 21-03-27
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Recently, with the development of sensor and VR technology, there are increasing opportunities to replace objects in real space with 3DCG models and handle them in virtual space. Therefore, one of the important technologies is posture estimation of 3D models. Posture estimation is to calculate the parameters required for deformation from the state of deformation of the 3D model. In this research, we describe the method of posture estimation by 3D Thin-plate Spline, the automatic division method of 3D model as its improvement method, and the generation of the intermediate model of deformation by deformation parameters.
View full abstract
-
Yoshiakira NAKAMURA, Tomonori IZUMI
Session ID: 21-03-28
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
The manuscript reports a trial of localization for a small robot vehicle supported by environmental cameras. We aim to improve the accuracy of localization of the robot utilizing images from multiple cameras installed within target area permanently, or temporally for setup or maintenance. We place webcams surrounding a miniature road course, calculate the position of a target marker, and evaluate the accuracy. OpenCV library is used to get the intrinsic parameters of the camera and to calibrate the distortion. We calculate the position from the camera images by using a simple gradient descent method. An experiment shows errors of 1.9cm and 1.7cm in average for 2 and 4 cameras, respectively
View full abstract
-
Tamie TANIGUCHI, Makoto J. HIRAYAMA
Session ID: 21-03-29
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In 2020, with the spread of the new coronavirus, online life has become more common, and the use of web conferencing systems has increased. To solve the inconvenience of being online, we have developed (1) a response analyzer and (2) a time management extension for the Google Meet web conferencing system. (1) allows teachers to immediately visualize students' responses to improve their classes. (2) reduces the burden of chairpersons in online conference presentations. In (1), we were able to display the feedback values on the buttons in an easy-to-understand manner. In addition, it was easy to see how many people responded and at what timing in the message field. In (2), we did not have to prepare timers and bells separately but could complete the process with Google Meet.
View full abstract
-
Shota Tsutsui, Junji Yamato, Jun Ohya, Ryo Sajima
Session ID: 21-03-30
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
The authors aim at achieving an autonomous robot monitors undeveloped woodland areas near a village, as they are under-managed due to aging and other factors. In this paper, for segmenting nighttime forest images by a deep learning based method, we investigate the feasibility of using Self-training to reduce the annotation cost of nighttime forest images and the effectiveness of the deep learning based segmentation method that uses the annotation. Our proposed method divides the dataset into labeled data and unlabeled data, and the unlabeled data are labeled by the threshold for the confidence level of the unlabeled data. As a result of comparing the segmentation results by supervised deep learning based segmentation using only labeled data and by the self-training based segmentation, even if the numbers of labeled and unlabeled data are small and large, respectively, we obtained promising results that demonstrate the validity of our proposed method.
View full abstract
-
Haruka IDE, Shuntaro AOTAKE, Hiroyuki OGATA, Jun OHYA, Takuya OHTANI, ...
Session ID: 21-03-31
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Under the Synecoculture environment, in which various plants are raised in mixed and dense vegetation, automatic maintenance of the field is difficult because of difficulties in separating each harvest. In this project, the situation in which one plant dominates the other plants is called “dominant situation”, and such dominant plants are to be cut. So, in this paper, we propose a method for detecting dominant plants from RGB images using deep learning. First, we partition the original image into small blocks. We perform VGG16 for each small block to predict the number of plants. If the number of the small blocks in each of which the number of plants is less than two exceeds the threshold, the original image is judged as a candidate of “dominant situation”. If the original image is judged as the candidate, similarity between dominant small blocks is computed using AKAZE, and if the similarity is high, the small blocks are judged to be in dominant situation. Experimental results show that high accuracies for estimating dominant situations are achieved.
View full abstract
-
Kazuhiro AKIZAWA, Yota YAMAMOTO, Yukinobu TANIGUCHI
Session ID: 21-03-32
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
We aim to develop a video-based dairy cow monitoring system that tracks cows in a whole barn using multiple cameras on the ceiling which share the field of views. In this system, the lens distortion of the cameras and the low frame rate of the videos lead to decrease the tracking accuracy. In this study, we propose a method that tracks cows across multiple cameras by merging the regions of the same individual cows detected in cameras sharing field of view in the world coordinate system. In addition, we improve the existing tracking algorithm to suppress the interruption of trajectory in low frame rate video. The experimental results suggest that our method improves the tracking accuracy.
View full abstract
-
Mamoru MATSUNAGA, Yota YAMAMOTO, Yukinobu TANIGUCHI
Session ID: 21-03-33
Published: 2022
Released on J-STAGE: March 31, 2023
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In the dairy industry, it is important to monitor the health of each dairy cow and to detect disease and estrus early. We have proposed a system based on image recognition to identify dairy cows by matching a database of multi-view RGB images generated from the 3D models and an input image of RGB images taken by ceiling cameras installed in the barn. This paper proposes a method to improve the matching accuracy by reducing the degradation of accuracy due to the difference in the generation process (domain) of database and input images. To further improve the accuracy, we also propose a method of negative mining to obtain efficient combinations of quadruplets.
View full abstract
