Skip to main content

kazuhiko sumi

Followers

0

Following

3

Public Views

Interests

Uploads

Papers by kazuhiko sumi

Optimization of Correlation Function and Sub-Pixel Estimation Method on Block Matching

In block matching based stereo, if we need precise depth, sub-pixel estimation of best matching l... more In block matching based stereo, if we need precise depth, sub-pixel estimation of best matching location is required. So far, many researches have been done on this sub-pixel estimation, but nothing is working well for real images. In this paper, I explained that the systematic error in sub-pixel estimation is generated by the mismatch of a correlation function and a fitting function, and I proved that the shape of correlation function is deformed from the ideal theoretical shape because of existing residual error of correlation function which is resulted in stereo calibration and digitization on block matching using real images. I proposed to model real correlation function by a linear combination of quadric and absolute difference. Applying this model, the systematic sub-pixel estimation error is mitigated, and it is effective also to existing the residual error of correlation function. The validity of this model was shown by applying to the simulation including some error and the...

Study on Template Protection for Biometric Authentication 鷲見和彦

This manuscript have been printed and distributed without reviewing and editing as received from ... more This manuscript have been printed and distributed without reviewing and editing as received from the authors: posting the manuscript to SCIS 2005 does not prevent future submission to any journals or conferences with proceedings. SCIS 2005 The 2005 Symposium on Cryptography and Information Security Maiko Kobe, Japan, Jan.25-28, 2005 The Institute of Electronics, Information and Communication Engineers

Self-Supervised Learning for Multi-View Stereo

Journal of the Japan Society for Precision Engineering, 2020

Automatic Recipe Recommendation from Food Flyers

Journal of the Japan Society for Precision Engineering, 2019

Speech-to-Gesture Generation Using Bi-Directional LSTM Network

Transactions of the Japanese Society for Artificial Intelligence, 2019

We present a novel framework for automatic speech-driven natural gesture motion generation. The p... more We present a novel framework for automatic speech-driven natural gesture motion generation. The proposed method consists of two steps. First, based on Bi-Directional LSTM Network, our deep network learns speech-gesture relationships with both forward and backward consistencies for a long period of time. The network regresses full 3D skeletal pose of a human from perceptual features extracted from the input audio in each time step. Second, we apply combined temporal filters to smooth out generated pose sequences. We utilize a speech-gesture dataset recorded with a headset and a marker-based motion capture to train our network. We evaluate different acoustic features, network architectures, and temporal filters in order to validate the effectiveness of the proposed approach. We also conduct a subjective evaluation and compare our approach against real human gestures. The subjective evaluation result shows that our generated gestures are comparable to "original" human gestures and are significantly better than "mismatched" human gestures taken from a different utterance in the scale of naturalness.

3D Digitization of a Hand-Held Object with a Wearable Vision Sensor

Lecture Notes in Computer Science, 2004

It is a common human behavior to hold a small object of interest and to manipulate it for observa... more It is a common human behavior to hold a small object of interest and to manipulate it for observation. A computer system, symbiotic with a human, should recognize the object and the intention of the human while the object is manipulated. To realize autonomous recognition of a hand-held object as well as to study human behavior of manipulation and observation, we have developed a wearable vision sensor, which has similar sight to a human who wears the sensor. In this paper, we show the following results obtained with the wearable vision sensor. First, we analyzed human manipulation for observation and related it with acquirable visual information. Second, we proposed dynamic space carving for 3D shape extraction of a static object occluded by a dynamic object moving around the static object. Finally, we showed that texture and silhouette information can be integrated in vacant space and the integration improves the efficiency of dynamic space carving.

Development of Road Surface Condition Detection System using Near-infrared Light- Absorption and Polarization Characteristics of Water

Transactions of the Society of Instrument and Control Engineers, 2010

多電極円形位置検出素子を用いた絶対回転角度計測

IEEJ Transactions on Industry Applications, 2009

Study on Synthetic Face Database for Performance Evaluation

Lecture Notes in Computer Science, 2005

We have analyzed the vulnerability and threat of the biometric evaluation database and proposed t... more We have analyzed the vulnerability and threat of the biometric evaluation database and proposed the method to generate a synthetic database from a real database. Our method is characterized by finding nearest neighbor triples or pairs in the feature space of biometric samples, and by crossing over those triples and pairs to generate synthetic samples. The advantages of our method is that we can keep the statistical distribution of the original database, thus, the evaluation result is expected to be the same as original real database. The proposed database, which does not have privacy problem, can be circulated freely among biometric vendors and testers. We have implemented this idea on a face image database using active appearance model. The synthesized image database has the same distance distribution with the original database, which suggests it will deriver the same accuracy with the original one.

非接触式開閉極動作行程センサの開発

IEEJ Transactions on Fundamentals and Materials, 2009

Biometrics Authentication and Application to Access Control Systems

The Journal of The Institute of Electrical Engineers of Japan, 2009

ＩＴＶ画像処理システムによるパルス溶接のサンプル値制御

QUARTERLY JOURNAL OF THE JAPAN WELDING SOCIETY, 1985

Frontiers of Image Processing Technologies and Applications. Application to ITS

The Journal of the Institute of Electrical Engineers of Japan

Towards Detailed 3D Modeling: Mesh Super-Resolution via Deformation

2020 Joint 9th International Conference on Informatics, Electronics & Vision (ICIEV) and 2020 4th International Conference on Imaging, Vision & Pattern Recognition (icIVPR), 2020

This paper presents a method for constructing a detailed mesh model by deforming a coarse mesh mo... more This paper presents a method for constructing a detailed mesh model by deforming a coarse mesh model using surface normals estimated from images. Mesh representation is a 3D data structure that can represent a detailed shape with fewer parameters compared with other data structures. Multi-view stereo (MVS) algorithms are suitable for reconstructing a detailed mesh model from images captured from different viewpoints; however, the computation time increases rapidly as the resolution and the number of images increase. To overcome this problem, we propose a novel approach to deform a coarse mesh model. We first reconstruct a coarse mesh model by applying the Delaunay triangulation to a coarse point cloud obtained through structure-from-motion. The resulting model is deformed based on the changes in the surface normals estimated from images. In our experiments, we demonstrate that our method constructs detailed 3D models faster than the MVS with a single model dataset and general scene ...

Optimization of Correlation Function and Sub-Pixel Estimation Method on Block Matching

In block matching based stereo, if we need precise depth, sub-pixel estimation of best matching l... more In block matching based stereo, if we need precise depth, sub-pixel estimation of best matching location is required. So far, many researches have been done on this sub-pixel estimation, but nothing is working well for real images. In this paper, I explained that the systematic error in sub-pixel estimation is generated by the mismatch of a correlation function and a fitting function, and I proved that the shape of correlation function is deformed from the ideal theoretical shape because of existing residual error of correlation function which is resulted in stereo calibration and digitization on block matching using real images. I proposed to model real correlation function by a linear combination of quadric and absolute difference. Applying this model, the systematic sub-pixel estimation error is mitigated, and it is effective also to existing the residual error of correlation function. The validity of this model was shown by applying to the simulation including some error and the...

Study on Template Protection for Biometric Authentication 鷲見和彦

This manuscript have been printed and distributed without reviewing and editing as received from ... more This manuscript have been printed and distributed without reviewing and editing as received from the authors: posting the manuscript to SCIS 2005 does not prevent future submission to any journals or conferences with proceedings. SCIS 2005 The 2005 Symposium on Cryptography and Information Security Maiko Kobe, Japan, Jan.25-28, 2005 The Institute of Electronics, Information and Communication Engineers

Self-Supervised Learning for Multi-View Stereo

Journal of the Japan Society for Precision Engineering, 2020

Automatic Recipe Recommendation from Food Flyers

Journal of the Japan Society for Precision Engineering, 2019

Speech-to-Gesture Generation Using Bi-Directional LSTM Network

Transactions of the Japanese Society for Artificial Intelligence, 2019

We present a novel framework for automatic speech-driven natural gesture motion generation. The p... more We present a novel framework for automatic speech-driven natural gesture motion generation. The proposed method consists of two steps. First, based on Bi-Directional LSTM Network, our deep network learns speech-gesture relationships with both forward and backward consistencies for a long period of time. The network regresses full 3D skeletal pose of a human from perceptual features extracted from the input audio in each time step. Second, we apply combined temporal filters to smooth out generated pose sequences. We utilize a speech-gesture dataset recorded with a headset and a marker-based motion capture to train our network. We evaluate different acoustic features, network architectures, and temporal filters in order to validate the effectiveness of the proposed approach. We also conduct a subjective evaluation and compare our approach against real human gestures. The subjective evaluation result shows that our generated gestures are comparable to "original" human gestures and are significantly better than "mismatched" human gestures taken from a different utterance in the scale of naturalness.

3D Digitization of a Hand-Held Object with a Wearable Vision Sensor

Lecture Notes in Computer Science, 2004

It is a common human behavior to hold a small object of interest and to manipulate it for observa... more It is a common human behavior to hold a small object of interest and to manipulate it for observation. A computer system, symbiotic with a human, should recognize the object and the intention of the human while the object is manipulated. To realize autonomous recognition of a hand-held object as well as to study human behavior of manipulation and observation, we have developed a wearable vision sensor, which has similar sight to a human who wears the sensor. In this paper, we show the following results obtained with the wearable vision sensor. First, we analyzed human manipulation for observation and related it with acquirable visual information. Second, we proposed dynamic space carving for 3D shape extraction of a static object occluded by a dynamic object moving around the static object. Finally, we showed that texture and silhouette information can be integrated in vacant space and the integration improves the efficiency of dynamic space carving.

Development of Road Surface Condition Detection System using Near-infrared Light- Absorption and Polarization Characteristics of Water

Transactions of the Society of Instrument and Control Engineers, 2010

多電極円形位置検出素子を用いた絶対回転角度計測

IEEJ Transactions on Industry Applications, 2009

Study on Synthetic Face Database for Performance Evaluation

Lecture Notes in Computer Science, 2005

We have analyzed the vulnerability and threat of the biometric evaluation database and proposed t... more We have analyzed the vulnerability and threat of the biometric evaluation database and proposed the method to generate a synthetic database from a real database. Our method is characterized by finding nearest neighbor triples or pairs in the feature space of biometric samples, and by crossing over those triples and pairs to generate synthetic samples. The advantages of our method is that we can keep the statistical distribution of the original database, thus, the evaluation result is expected to be the same as original real database. The proposed database, which does not have privacy problem, can be circulated freely among biometric vendors and testers. We have implemented this idea on a face image database using active appearance model. The synthesized image database has the same distance distribution with the original database, which suggests it will deriver the same accuracy with the original one.

非接触式開閉極動作行程センサの開発

IEEJ Transactions on Fundamentals and Materials, 2009

Biometrics Authentication and Application to Access Control Systems

The Journal of The Institute of Electrical Engineers of Japan, 2009

ＩＴＶ画像処理システムによるパルス溶接のサンプル値制御

QUARTERLY JOURNAL OF THE JAPAN WELDING SOCIETY, 1985

Frontiers of Image Processing Technologies and Applications. Application to ITS

The Journal of the Institute of Electrical Engineers of Japan

Towards Detailed 3D Modeling: Mesh Super-Resolution via Deformation

2020 Joint 9th International Conference on Informatics, Electronics & Vision (ICIEV) and 2020 4th International Conference on Imaging, Vision & Pattern Recognition (icIVPR), 2020

This paper presents a method for constructing a detailed mesh model by deforming a coarse mesh mo... more This paper presents a method for constructing a detailed mesh model by deforming a coarse mesh model using surface normals estimated from images. Mesh representation is a 3D data structure that can represent a detailed shape with fewer parameters compared with other data structures. Multi-view stereo (MVS) algorithms are suitable for reconstructing a detailed mesh model from images captured from different viewpoints; however, the computation time increases rapidly as the resolution and the number of images increase. To overcome this problem, we propose a novel approach to deform a coarse mesh model. We first reconstruct a coarse mesh model by applying the Delaunay triangulation to a coarse point cloud obtained through structure-from-motion. The resulting model is deformed based on the changes in the surface normals estimated from images. In our experiments, we demonstrate that our method constructs detailed 3D models faster than the MVS with a single model dataset and general scene ...