論文

Occlusion-free video generation is challenging due to surgeons’ obstructions in the camera field of view. Prior work has addressed this issue by installing multiple cameras on a surgical light, hoping some cameras will observe the surgical field with less occlusion. However, this special camera setup poses a new imaging challenge since camera configurations can change every time surgeons move the light, and manual image alignment is required. This paper proposes an algorithm to automate this alignment task. The proposed method detects frames where the lighting system moves, realigns them, and selects the camera with the least occlusion. This algorithm results in a stabilized video with less occlusion. Quantitative results show that our method outperforms conventional approaches. A user study involving medical doctors also confirmed the superiority of our method.

Identifying individual humpback whales by photographs of their tails is valuable for understanding the ecology of wild whales. We have about 10,000 photos of 1,850 identified whales taken in the sea area around Okinawa over a 30-year period. The identification process on this large scale of numbers is difficult not only for the human eye but also for machine vision, as the numbers of photographs per individual whale are very low. About 30% of the whales have only a single photograph, and 80% have fewer than five. In addition, the shapes of the tails and the black and white patterns on them are vague, and these change readily with the whale’s slightest movement and changing photo-shooting conditions. We propose a practical method for identifying a humpback whale by accurate segmentation of the fluke region using a combination of deep neural networking and GrabCut. Then useful features for identifying each individual whale are extracted by both histograms of image features and wavelet transform of the trailing edge. The test results for 323 photos show the correct individuals are ranked within the top 30 for 89% of the photos, and at the same time for 76% of photos ranked at the top.

Novel view synthesis from images, for example, with 3D Gaussian splatting, has made great progress. Rendering fidelity and speed are now ready even for demanding virtual reality applications. However, the problem of assisting humans in collecting the input images for these rendering algorithms has received much less attention. High-quality view synthesis requires uniform and dense view sampling. Unfortunately, these requirements are not easily addressed by human camera operators, who are in a hurry, impatient, or lack understanding of the scene structure and the photographic process. Existing approaches to guide humans during image acquisition concentrate on single objects or neglect view-dependent material characteristics. We propose a novel situated visualization technique for scanning at multiple scales. During the scanning of a scene, our method identifies important objects that need extended image coverage to properly represent view-dependent appearance. To this end, we leverage semantic segmentation and category identification, ranked by a vision-language model. Spherical proxies are generated around highly ranked objects to guide the user during scanning. Our results show superior performance in real scenes compared to conventional view sampling strategies.

Food picking is trivial for humans but not for robots, as foods are fragile. Presetting foods' physical properties does not help robots much due to the objects' inter- and intra-category diversity. A recent study proved that learning-based fracture anticipation with tactile sensors could overcome this problem; however, the method trains the model for each food to deal with intra-category differences, and tuning robots for each food leads to an undesirable amount of food consumption. This study proposes a novel framework for learning food-picking tasks without consuming foods. The key idea is to leverage the object-breaking experiences of several reusable fragile objects instead of consuming real foods while making the picking ability object-invariant with domain generalization (DG). In real-robot experiments, we trained a model with reusable objects (toy blocks, ping-pong balls, and jellies), selected based on the three common fracture types (crack, rupture, and crush). We then tested the model with four real food objects (tofu, bananas, potato chips, and tomatoes). The results showed that the proposed combination of reusable objects' breaking experiences and DG is effective for the food-picking task.

How much can we infer about human shape using an event camera that only detects the pixel position where the luminance changed and its timestamp? This neuromorphic vision technology captures changes in pixel values at ultra-high speeds, regardless of the variations in environmental lighting brightness. Existing methods for human mesh recovery (HMR) from event data need to utilize intensity images captured with a generic frame-based camera, rendering them vulnerable to low-light conditions, energy/memory constraints, and privacy issues. In contrast, we explore the potential of solely utilizing event data to alleviate these issues and ascertain whether it offers adequate cues for HMR, as illustrated in Fig. 1. This is a quite challenging task due to the substantially limited information ensuing from the absence of intensity images. To this end, we propose EventPointMesh, a framework which treats event data as a three-dimensional (3D) spatio-temporal point cloud for reconstructing the human mesh. By employing a coarse-to-fine pose feature extraction strategy, we extract both global features and local features. The local features are derived by processing the spatio-temporally dispersed event points into groups associated with individual body segments. This combination of global and local features allows the framework to achieve a more accurate HMR, capturing subtle differences in human movements. Experiments demonstrate that our method with only sparse event data outperforms baseline methods.

Machine vision systems play a pivotal role in streamlining manufacturing processes, notably in quality control through automatic in-line visual inspections. A common practice for inspecting parts, components, and final products is to use a master part benchmark for quality comparison. However, challenges arise when objects enter inspection points in unintended orientations. This misalignment potentially leads to erroneous decisions by automated systems, resulting in additional checkpoints or wastage affecting the production rate. To tackle this issue, we propose a visual inspection pipeline that leverages recent machine learning-based approaches to compare the inspection target and a master part virtually oriented to the same perspective. Specifically, we suggest combining 3D Gaussian Splatting and DUSt3R as a practical solution. Our approach demonstrates its efficacy in real-world scenarios through testing on three mock parts and a real industrial component.

Video analysis has become an essential aspect of net sports, such as badminton. Accurately predicting the future trajectory of balls and shuttlecocks can significantly benefit players by enhancing their performance and enabling them to devise effective game strategies. This paper aims to analyze data to provide players with an advantage in the fast-paced rallies of badminton matches. The paper delves into the innovative task of predicting future shuttlecock trajectories in badminton match videos and presents a method that takes into account both the shuttlecock position and the positions and postures of the players. In the experiments, players were extracted from the match video, their postures were analyzed, and a time-series model was trained. The results indicate that the proposed method improved accuracy by 13% compared to methods that solely used shuttlecock position information as input, and by 8.4% compared to methods that employed both shuttlecock and player position information as input.

Multi-Object Tracking (MOT) plays a critical role in analyzing player behavior from videos, enabling performance evaluation. Current MOT methods are often evaluated using publicly available datasets. However, most of these focus on everyday scenarios such as pedestrian tracking or are tailored to specific sports, including soccer and basketball. Despite the inherent challenges of tracking players in American football, such as frequent occlusion and physical contact, no standardized dataset has been publicly available, making fair comparisons between methods difficult. To address this gap, we constructed the first dedicated detection and tracking dataset for the American football players and conducted a comparative evaluation of various detection and tracking methods. Our results demonstrate that accurate detection and tracking can be achieved even in crowded scenarios. Fine-tuning detection models improved performance over pre-trained models. Furthermore, when these fine-tuned detectors and re-identification models were integrated into tracking systems, we observed notable improvements in tracking accuracy compared to existing approaches. This work thus enables robust detection and tracking of American football players in challenging, high-density scenarios previously underserved by conventional methods.

Ball spin estimation in sports is important for analyzing the game. Since spin is generally too fast to be captured by a conventional camera, a high-speed camera is often used to capture images of the ball and estimate its spin. However, since a high-speed camera is not robust to changes in the lighting conditions, it is difficult to estimate spin in some environments. To solve these problems, this paper proposes a new method for ball spin estimation using an event camera. An event camera is a sensor inspired by the visual system of animals, which outputs the brightness changes in a scene. Event cameras have advantages such as high temporal resolution and high dynamic range, and can accurately capture the motion of a fast-spinning ball in various lighting conditions. Experimental results in a synthesized dataset showed that the proposed method can stably estimate spin up to 500 rps. It is also confirmed that the proposed method can estimate spin in the data obtained from actual sports games.

Motion blur can significantly reduce the quality of images, and researchers have developed various algorithms to address this issue. One common approach to deblurring is to use deconvolution to cancel out the blur effect, but this method is limited by the difficulty of accurately estimating blur kernels from blurred images. This is because the motion causing the blur is often complex and nonlinear. In this paper, a new method for estimating blur kernels is proposed. This method uses an event camera, which captures high-temporal-resolution data on pixel luminance changes, along with a conventional camera to capture the input blurred image. By analyzing the event data stream, the proposed method estimates the 2D motion of the blurred image at short intervals during the exposure time, and integrates this information to estimate a variety of complex blur motions. With the estimated blur kernel, the input blurred image can be deblurred using deconvolution. The proposed method does not rely on machine learning and therefore can restore blurry images without depending on the quality and quantity of training data. Experimental results show that the proposed method can estimate blur kernels even for images blurred by complex camera motions, outperforming conventional methods. Overall, this paper presents a promising approach to motion deblurring that could have practical applications in a range of fields.

When a camera travels across a 3D world, only a fraction of pixel value changes; an event-based camera observes the change as sparse events. How can we utilize sparse events for efficient recovery of the camera pose? We show that we can recover the camera pose by minimizing the error between sparse events and the temporal gradient of the scene represented as a neural radiance field (NeRF). To enable the computation of the temporal gradient of the scene, we augment NeRF's camera pose as a time function. When the input pose to the NeRF coincides with the actual pose, the output of the temporal gradient of NeRF equals the observed intensity changes on the event's points. Using this principle, we propose an event-based camera pose tracking framework called TeGRA which realizes the pose update by using the sparse event's observation. To the best of our knowledge, this is the first camera pose estimation algorithm using the scene's implicit representation and the sparse intensity change from events.

Ball detection in sports, particularly in fast-paced games like volleyball, where the ball is constantly in high motion, presents a significant challenge for game analysis and automated sports broadcasting. Conventional camera-based ball detection faces issues, such as motion blur, in high-speed ball movement scenes. To address these challenges, we propose a deep learning-based method for detecting balls using event cameras. Event cameras, also known as dynamic vision sensors, operate differently from traditional cameras. Instead of capturing frames at fixed intervals, they record individual pixel-level luminance changes, referred to as events. This unique feature enables event cameras to provide precise temporal information with low latency. Our proposed method transforms sparse events into an image format, enabling the use of current deep-learning architectures for object detection. Given the limited amount of events available for training an object detector, we generate synthetic events from RGB frames. This approach reduces the need for extensive annotation and ensures sufficient data availability. Experimental results confirm that our proposed method can detect balls that are undetectable in RGB frames and outperform existing methods that utilize event-based ball detection. Moreover, we conducted tests to verify our method's ability to detect balls in real events, not just synthetic ones. These results demonstrate that our proposed method opens up new possibilities in sports ball detection.

Training and education have become increasingly crucial in obtaining new skills in a variety of fields, especially in assembly and disassembly operations. The main issue in mechanical engineering, particularly in the assembly department, was that automobile engine components assembly was found to be complicated and challenging to assemble using an existing method, where they only rely on a video-based method. The purpose of this paper is to create interactive Virtual Reality (VR) and Augmented Reality (AR) applications that allow users to efficiently assist and complete the assembly tasks. In this work, the authors designed and developed a fully immersive VR application using an HTC Vive headset and two AR applications (marker-less AR application and marker-based AR application) using EPSON MOVERIO BT-300 (AR Smart Glasses). Fourteen engineering students from Universiti Putra Malaysia were selected for the experiment. They were divided into four groups: video-based group, VR-based group, marker-less AR group, and marker-based AR group. They are required to complete all four experiments (video-based experiment, VR-based experiment, marker-less AR experiment, and marker-based AR experiment). The results showed that the marker-less AR application is the best impressive method (37% better), the VR application is the second impressive method (23% better) followed by the marker-based AR application is the third impressive method (3% better) compared to the existing video-based guideline. Therefore, the students favored AR and VR applications rather than the existing method to be used in automobile engine assembly tasks.

Only virtual hands and feet move synchronously with an observer's movement, inducing body ownership of an invisible body between them. However, it is unclear whether body ownership is also induced when the size of the invisible body is changed. In this study, we investigated whether body ownership is induced in large or small invisible bodies and whether size perception changes with the size of the invisible body. The results showed that body ownership was induced even if the size of the invisible body was changed, but the size perception did not change.

Egocentric open-surgery videos capture rich, fine-grained details essential for accurately modelling surgical procedures and human behaviour in the operating room. A detailed, pixel-level understanding of hands and surgical tools is crucial for interpreting a surgeon action and intention. We introduce EgoSurgery-HTS, a new dataset with pixel-wise annotations and a benchmark suite for segmenting surgical tools, hands, and interacting tools in egocentric open-surgery videos. Specifically, we provide a labelled dataset for (1) tool instance segmentation of 14 distinct surgical tools, (2) hand instance segmentation, and (3) hand–tool segmentation to label main operating hands and the tools they manipulate. Using EgoSurgery-HTS, we conduct extensive evaluations of state-of-the-art segmentation methods and demonstrate significant improvements in the accuracy of hand and hand–tool segmentation in egocentric open-surgery videos compared to existing datasets. The dataset will be released upon acceptance.

High-speed phenomena, which are important in fields such as sports and fluid dynamics, are often challenging for the human visual system to perceive precisely. To better understand these high-speed phenomena, it is important to present users the information intuitively and precisely without time delays. Existing visualization techniques, such as high-speed cameras and stroboscopes, are effective for capturing high-speed events but are not well-suited for interactive visualization purposes.In this study, we propose an interactive visualization system of high-speed events using an event camera. The system precisely captures high-speed phenomena and allows users to replay important scenes instantly with a single key press. Through a user study involving 38 participants, we confirmed that the event camera system enhances human visual perception. Furthermore, a case study demonstrated that the interactive design enables users to interpret high-speed phenomena more intuitively. These findings suggest that our system can effectively augment users' visual perception.

Most robots implicitly assume that the road surface on which they move is flat, without differences in level. Detecting differences in level on roads contributes to robots moving safely without stacking and falling. Although there are some studies on detecting differences in level in RGB or RGB-D images, directly finding only differences in level on roads is difficult due to the abundance and complexity of the types of differences in level on roads. This paper presents a new method for detecting differences in level from RGB-D images obtained by a modern smartphone equipped with a high-performance depth camera. First, we extract a part of differences in level on roads by finding the change of the normal vector in the contour of the detected plane. Then, a deep learning model trained on the dataset created by using the extracted image patches is used for detecting all the differences in level in outdoor images. To evaluate the effectiveness of the proposed method, quantitative and qualitative comparisons with existing methods were conducted. Further, the results from various inputs were qualitatively and quantitatively evaluated. As a result, we verified that the proposed method was able to detect all differences in level in an image, even in complex scenes where existing methods cannot detect.

Although recent deep learning-based inpainting techniques achieve excellent restoration quality, their stringent requirement for computational resources such as GPU/VRAM render them difficult to use in settings where high-end hardware is not available. We address this gap between research and real-world applications in our submission to the DREAMING challenge, which explores such a use case - diminished reality - where run time and GPU hardware are limited. Specifically, this paper proposes a method that divides a video optimally into sections of variable lengths for the downstream inpainting model. By maximizing the number of frames, i.e., spatio-temporal information, in each subvideo, inference using a SOTA model can often be performed without significant degradation to output quality. In addition, to expedite inference, we examine techniques that intelligently reduce the amount of input pixels, e.g., downsampling and cropping, while maintaining decent inpainting quality.

Event cameras, known for their high dynamic range, absence of motion blur, and low energy usage, have recently found a wide range of applications thanks to these attributes. In the past few years, the field of event-based 3D reconstruction saw remarkable progress, with the Neural Radiance Field (NeRF) based approach demonstrating photorealistic view synthesis results. However, the volume rendering paradigm of NeRF necessitates extensive training and rendering times. In this paper, we introduce Event Enhanced Gaussian Splatting (E2GS), a novel method that incorporates event data into Gaussian Splatting, which has recently made significant advances in the field of novel view synthesis. Our E2GS effectively utilizes both blurry images and event data, significantly improving image deblurring and producing high-quality novel view synthesis. Our comprehensive experiments on both synthetic and real-world datasets demonstrate our E2GS can generate visually appealing renderings while offering faster training and rendering speed (140 FPS). Our code is available at https://github.com/deguchihiroyuki/E2GS.

A crowd density forecasting task aims to predict how the crowd density map will change in the future from observed past crowd density maps. However, the past crowd density maps are often incomplete due to the miss-detection of pedestrians, and it is crucial to develop a robust crowd density forecasting model against the miss-detection. This paper presents a MAsked crowd density Completion frame-work for crowd density forecasting (CrowdMAC), which is simultaneously trained to forecast future crowd density maps from partially masked past crowd density maps (i.e., forecasting maps from past maps with miss-detection) while reconstructing the masked observation maps (i.e., imputing past maps with miss-detection). Additionally, we propose Temporal-Density-aware Masking (TDM), which non-uniformly masks tokens in the observed crowd density map, considering the sparsity of the crowd density maps and the informativeness of the subsequent frames for the forecasting task. Moreover, we introduce multitask masking to enhance training efficiency. In the experiments, CrowdMAC achieves state-of-the-art performance on seven large-scale datasets, including SDD, ETH-UCY, inD, JRDB, VSCrowd, FDST, and croHD. We also demonstrate the robustness of the proposed method against both synthetic and realistic miss-detections. The code is released at project page.