論文

© 2018 IEEE. We propose a novel three-dimensional (3D)-convolution method, cv3dconv, for detecting spatiotemporal features from videos. It reduces the number of sum-of-products of 3D convolution by thousands of times by assuming the constant moving velocity of the camera. We observed that a specific class of video sequences, such as those captured by an in-vehicle camera, can be well approximated with piece-wise linear movements of 2D features in the temporal dimension. Our principal finding is that the 3D kernel, represented by the constant-velocity, can be decomposed into a convolution of a 2D kernel representing the shapes and a 3D kernel representing the velocity. We derived the efficient recursive algorithm for this class of 3D convolution which is exceptionally suited for sparse data, and this parameterized decomposed representation imposes a structured regularization along the temporal direction. We experimentally verified the validity of our approximation using a controlled dataset, and we also showed the effectiveness of cv3dconv for the visual odometry estimation task using real event camera data captured in urban road scene.

© 2018 IEEE. We propose a novel 3-D convolution method, cv3dconv, for extracting spatiotemporal features from videos. It reduces the number of sum-of-products operations in 3-D convolution by thousands of times by assuming the constant moving velocity of the features. We observed that a specific class of video sequences, such as video captured by an in-vehicle camera, can be well approximated with piece-wise linear movements of 2-D features in a temporal dimension. Our principal finding is that a 3-D kernel, represented by constant velocity, can be decomposed into a convolution of a 2-D-shaped kernel and a 3-D-velocity kernel, which is parameterized using only two parameters. We derived an efficient recursive algorithm for this class of 3-D convolution, which is exceptionally suited for sparse spatiotemporal data, and this parameterized decomposed representation imposes a structured regularization along a temporal direction. We experimentally verified the validity of our approximation using a controlled dataset, and we also showed the effectiveness of the cv3dconv by adopting it for deep neural networks (DNNs) in visual odometry estimation task using publicly available event-based camera dataset captured in urban road scene. Our DNN architecture improves the estimation accuracy for about 30% compared with the existing states-of-the-arts architecture designed for event data.

© 2016 IEEE. Point cloud registration is a key problem for robotics and computer vision communities. This represents estimating a rigid transform which aligns one point cloud to another. Iterative closest point (ICP) is a well-known classical method for this problem, yet it generally achieves high alignment only when the source and template point cloud are mostly pre-aligned. If each point cloud is far away or contains a repeating structure, the registration often fails because of being fallen into a local minimum. Recently, inspired by PointNet, several deep learning-based methods have been developed. PointNetLK is a representative approach, which directly optimizes the distance of aggregated features using gradient method by Jacobian. In this paper, we propose a point cloud registration system based on deep learning: CorsNet. Since CorsNet concatenates the local features with the global features and regresses correspondences between point clouds, not directly pose or aggregated features, more useful information is integrated than the conventional approaches. For comparison, we also developed a novel deep learning approach (DirectNet) that directly regresses the pose between point clouds. Through our experiments, we show that CorsNet achieves higher accuracy than not only the classic ICP method, but also the recently proposed learning-based proposal PointNetLK and DirectNet, including on seen and unseen categories.

© 2020, Springer Nature Switzerland AG. Recording surgery in operating rooms is an essential task for education and evaluation of medical treatment. However, recording the desired targets, such as the surgery field, surgical tools, or doctor’s hands, is difficult because the targets are heavily occluded during surgery. We use a recording system in which multiple cameras are embedded in the surgical lamp, and we assume that at least one camera is recording the target without occlusion at any given time. As the embedded cameras obtain multiple video sequences, we address the task of selecting the camera with the best view of the surgery. Unlike the conventional method, which selects the camera based on the area size of the surgery field, we propose a deep neural network that predicts the camera selection probability from multiple video sequences by learning the supervision of the expert annotation. We created a dataset in which six different types of plastic surgery are recorded, and we provided the annotation of camera switching. Our experiments show that our approach successfully switched between cameras and outperformed three baseline methods.

© 2019. The copyright of this document resides with its authors. We present DetectFusion, an RGB-D SLAM system that runs in real time and can robustly handle semantically known and unknown objects that can move dynamically in the scene. Our system detects, segments and assigns semantic class labels to known objects in the scene, while tracking and reconstructing them even when they move independently in front of the monocular camera. In contrast to related work, we achieve real-time computational performance on semantic instance segmentation with a novel method combining 2D object detection and 3D geometric segmentation. In addition, we propose a method for detecting and segmenting the motion of semantically unknown objects, thus further improving the accuracy of camera tracking and map reconstruction. We show that our method performs on par or better than previous work in terms of localization and object reconstruction accuracy, while achieving about 20 fps even if the objects are segmented in each frame.

© 2020 IEEE. This paper presents a method for automatic detection of cervical myelopathy by using Leap Motion Sensor for sensing hand motion to apply a method called the 10-s grip and release test. Leap Motion data is analyzed for computing feature of hand motion, which is used for automatic detection of patient or healthy people by random forests algorithm for machine learning. For evaluating automatic detection performance, we collect real data of hand motion 54 of patient and healthy people, which is used for validating accuracy of detection.

© 2018 IEEE. Incrementally, building a 3D map in which object instances are semantically annotated has a wide range of applications, including scene understanding, human-robot interactions, and simultaneous localization and mapping extensions. Although researchers are developing efficient and accurate systems, these methods still face a critical issue: Real-time processing, because the task requires a series of heavy processing components, e.g., camera pose estimation, 3D map reconstruction, and especially recognition. In this paper, we propose a novel object-oriented semantic mapping approach aiming at overcoming such issues by introducing highly accurate object-oriented semantic scene reconstruction in real time. For high efficiency, the proposed method employs a fast and scalable object detection algorithm for exploiting semantic information from the incoming frames. These outputs are integrated into geometric regions of the 3D map, which are carried by the geometric-based incremental segmentation method. The strategy of assigning class probabilities to each segmented region, not each element (e.g., surfels and voxels), notably reduces the computational cost, as well as the memory footprint. In addition to efficiency, by geometrically segmenting the 3D map first, clear boundaries between objects appear. We complementarily improve the geometric-based segmentation results beyond the geometric only to the semantic-aware representation. We validate the proposed method's accuracy and computational efficiency through experiments in a common office scene.

© 2018 IEEE. This paper presents a method to estimate the 3D position of a center of mass (CoM) of a human body from a set of multi-view images. As a well-known fact, in sports, collections of CoM are important for analyzing the athletes' performance. Most conventional studies in CoM estimation require installing a measuring system (e.g., a force plate or optical motion capture system) or attaching sensors to the athlete. While such systems reliably estimate CoM, casual settings are preferable for simplifying preparations. To address this issue, the proposed method takes a vision-based approach that does not require specialized hardware and wearable devices. Our method calculates subject's CoM using voxels with body parts dependent weighting. This individual voxel reconstruction and voxel-wise weighting reflects the differences in each body shape, and are expected to contribute to higher performance in analysis. The results using real data demonstrated the performance of the proposed method were compared to force plate data, and provided a 3D CoM visualization in a dynamic scene.

© 2018 Copyright held by the owner/author(s). Publication rights licensed to ACM. The purpose of this study is sensing movements of 100-m runners from video that is publicly available, for example, Internet broadcasts. Normally, information that can be obtained from a video is limited to the number of steps and average stride length. However, our proposed method makes it possible to measure not only this information, but also time-scale information like every stride length and speed transition from the same input. Our proposed method can be divided into three steps. First, we generate a panoramic image of the 100-m track. By doing this, we can estimate where the runners are running in a frame at the 100-meter scale. Second, we detect whether the runner steps in the frame. For this process, we utilize the detected track lines and leg joint positions of runners. Finally, we project every steps to the overview image of the 100-m track to estimate the stride length at the 100-m scale. In the experiment part, we apply our method to various race videos. We evaluate the accuracy of our method via comparison with the data measured using typical methods. In addition, we evaluate the accuracy of estimation of the number of steps and show visualized runners’ steps and speed transitions.

© 2019 IEEE. Event cameras are bio-inspired vision sensors that mimic retinas to asynchronously report per-pixel intensity changes rather than outputting an actual intensity image at regular intervals. This new paradigm of image sensor offers significant potential advantages; namely, sparse and non-redundant data representation. Unfortunately, however, most of the existing artificial neural network architectures, such as a CNN, require dense synchronous input data, and therefore, cannot make use of the sparseness of the data. We propose EventNet, a neural network designed for real-time processing of asynchronous event streams in a recursive and event-wise manner. EventNet models dependence of the output on tens of thousands of causal events recursively using a novel temporal coding scheme. As a result, at inference time, our network operates in an event-wise manner that is realized with very few sum-of-the-product operations--look-up table and temporal feature aggregation--which enables processing of 1 mega or more events per second on standard CPU. In experiments using real data, we demonstrated the real-time performance and robustness of our framework.

This paper presents a method for synthesizing a stroboscopic image of a moving sports player from a hand-held camera sequence. This method has three steps: synthesis of background image, synthesis of stroboscopic image, and removal of player’s shadow. In synthesis of background image step, all input frames masked a bounding box of the player are stitched together to generate a background image. The player is extracted by an HOG-based people detector. In synthesis of stroboscopic image step, the background image, the input frame, and a mask of the player synthesize a stroboscopic image. In removal of shadow step, we remove the player’s shadow which negatively affects an analysis by using mean-shift. In our previous work, synthesis of background image has been timeconsuming. In this paper, by using the bounding box of the player detected by HOG and by subtracting the images for synthesizing a mask, computational speed and accuracy can be improved. These have contributed greatly to the improvement from the previous method. These are main improvements and novelty points from our previous method. In experiments, we confirmed the effectiveness of the proposed method, measured the player’s speed and stride length, and made a footprint image. The image sequence was captured under a simple condition that no other people were in the background and the person controlling the video camera was standing still, such like a motion parallax was not occurred. In addition, we applied the synthesis method to various scenes to confirm its versatility.

In this paper, we propose a method for recognition and pose estimation of primitive shapes from depth images for spatial augmented reality (SAR). To use SAR in our everyday life, technology to recognize and estimate the pose of projected objects in the room is necessary. However, it is not a simple task to recognize primitive shapes because of their low 3D feature values.

The authors propose a line-segment-based marker-less augmented reality (AR) framework that involves an on-site model-generation method and on-line camera tracking. In most conventional model-based marker-less AR frameworks, correspondences between the 3D model and the 2D frame for camera-pose estimation are obtained by feature-point matching. However, 3D models of the target scene are not always available, and feature points are not detected from texture-less objects. The authors' framework is based on a model-generation method with an RGB-D camera and model-based tracking using line segments, which can be detected even with only a few feature points. The camera pose of the input images can be estimated from the 2D-3D line-segment correspondences given by a line-segment feature descriptor. The experimental results show that the proposed framework can achieve AR when other point-based frameworks cannot. The authors also argue that their framework can generate a model and estimate camera pose more accurately than their previous study.

The illumination chromaticity is one of the important factor for appearance of objects in images. When we superimpose some objects on another scene to generate composite images, the resulting image may give us strangeness if the illumination chromaticity of foreground scene and background scene are different. In this paper, we propose a composite image synthesis method which can achieve illumination chromaticity consistency. Our goal is to generate a illumination chromaticity-matched composite image from two images which are taken under different illumination environment. We estimate chromaticity of two scenes based on segmented color regions. Subsequently illumination chromaticity matching is performed based on estimated chromaticity of two scenes. After the chromaticity matching process, we extract the target person by using the grab cut segmentation. The result image is obtained by superimposing the target person or object onto background scene. We demonstrate that our method can perform some different illumination chromaticity environment.

In this paper, we propose a method for camera pose estimation based on keypoints matching with images capturing the target scene from multiple viewpoints. For estimating camera pose from an image, 3D structure of the target scene is needed. Then a set of correspondences between 2D position in the image and 3D positions in the target scene provides the camera pose by solving Perspective-n-Point problem. Image keypoints are used as these positions, but finding corresponding 3D positions of the 3D structure is not easy problem even if a 3D textured model of the target scene is given. This is because the appearance of each 3D position is not always the same. In order to solve this problem, we adopt real images as a source of keypoints contained in the database. Real image represents accurate appearance for each 3D position, so keypoint matching with the input images can be more accurate. We constructed database based on the keypoint using the local image feature derived from real images and evaluated the accuracy of camera pose estimation. The result showed that the database constructed by our method is effective to accurate camera tracking.

A method for synthesizing the stroboscopic image of a moving sports player from a hand-held camera sequence is proposed. In this method, a player is extracted from each frame of an input video sequence using a HOG-based people detector. After removing the bounding box of the extracted player from each frame, all frames are stitched together to generate a background image without the player. The player's area is then overlaid onto the stitched background image and a stroboscopic image of the player is synthesized. By using the bounding box of the player detected by HOG and by subtracting the images, computational speed and accuracy can be improved compared with the conventional method that segments the player's region. In addition to the stroboscopic image synthesis, we also remove the player's shadow to improve the appearance of the resultant stroboscopic image. Experimental results demonstrate the effectiveness of the proposed method.

Relighting techniques can achieve the photometric consistency in synthesizing a composite images. Relighting generally need the object’s shape and illumination environment. Recent research demonstrates possibilities of relighting for an unknown shape object or the relighting the object under unknown illumination environment. However, achieving both tasks are still challenging issue. In this paper, we propose a relighting method for an unknown shape object captured under unknown illumination environment by using an RGB-D camera. The relighted object can be rendered from pixel intensity, surface albedo and shape of the object and illumination environment. The pixel intensity and the shape of the object can simultaneously be obtained from an RGB-D camera. Then surface albedo and illumination environment are iteratively estimated from the pixel intensity and shape of the object.We demonstrate that our method can perform relighting for a dynamic shape object captured under unknown illumination using an RGB-D camera.

We propose a method of ego-motion estimation for a self-driving car on which we installed cameras with non-overlapping views. By finding corresponding points between the multi-camera images, we aim to enhance the accuracy of the ego-motion estimation. However since the viewing directions are very different from one camera to the other, a conventional algorithm such as SURF cannot detect a sufficient number of correspondences. Additionally in the case where cameras have low frame rate and the vehicle has high speed, the scene changes might be too big to find correspondences between the same camera images. We propose a novel matching algorithm by warping feature patches detected in different cameras based on urban 3D structure. We assume that detected features exist on the surface of buildings or roads and the patch around the feature is planar. Based on this assumption, we can warp the patches so that the feature descriptors are similar for the corresponding feature points. We apply Bundle Adjustment to the found correspondences to optimize the odometry. The result shows higher estimation accuracy when compared to other matching methods.

We propose a method of ego-motion estimation for a self-driving car on which we installed cameras with non-overlapping views. By finding corresponding points between the multi-camera images, we aim to enhance the accuracy of the ego-motion estimation. However since the viewing directions are very different from one camera to the other, a conventional algorithm such as SURF cannot detect a sufficient number of correspondences. Additionally in the case where cameras have low frame rate and the vehicle has high speed, the scene changes might be too big to find correspondences between the same camera images. We propose a novel matching algorithm by warping feature patches detected in different cameras based on urban 3D structure. We assume that detected features exist on the surface of buildings or roads and the patch around the feature is planar. Based on this assumption, we can warp the patches so that the feature descriptors are similar for the corresponding feature points. We apply Bundle Adjustment to the found correspondences to optimize the odometry. The result shows higher estimation accuracy when compared to other matching methods.