A5091227949- Human Pose and Action Recognition
- Multimodal Machine Learning Applications
- Advanced Neural Network Applications
- Video Analysis and Summarization
- Hand Gesture Recognition Systems
- Image Enhancement Techniques
- Advanced Image and Video Retrieval Techniques
- Advanced Vision and Imaging
- Visual Attention and Saliency Detection
- Robot Manipulation and Learning
- Generative Adversarial Networks and Image Synthesis
- Railway Systems and Energy Efficiency
- Domain Adaptation and Few-Shot Learning
- Electrical Contact Performance and Analysis
- Text and Document Classification Technologies
- Engineering Applied Research
- Video Surveillance and Tracking Methods
- Anomaly Detection Techniques and Applications
- Computer Graphics and Visualization Techniques
- Image Retrieval and Classification Techniques
- Topic Modeling
- Innovation in Digital Healthcare Systems
- CCD and CMOS Imaging Sensors
- Electric and Hybrid Vehicle Technologies
- Robotics and Sensor-Based Localization
The University of Tokyo
2015-2025
Tokyo University of Science
2019
Tokyo University of Information Sciences
2015-2018
East Japan Railway (Japan)
2005-2014
Aims Community College
2005
University of Washington
1983
Can we detect common objects in a variety of image domains without instance-level annotations? In this paper, present framework for novel task, cross-domain weakly supervised object detection, which addresses question. For have access to images with annotations source domain (e.g., natural image) and image-level target watercolor). addition, the classes be detected are all or subset those domain. Starting from fully detector, is pre-trained on domain, propose two-step progressive adaptation...
This article tackles a new problem setting: reinforcement learning with pixel-wise rewards (pixelRL) for image processing. After the introduction of deep Q-network, RL has been achieving great success. However, applications (RL) processing are still limited. Therefore, we extend to pixelRL various applications. In pixelRL, each pixel an agent, and agent changes value by taking action. We also propose effective method that significantly improves performance considering not only future states...
East Japan Railway Company (JR East) has developed the catenary and storage battery hybrid train system using a test car for purpose of through operation service between electrified section non-electrified decreasing environmental impact diesel trains operating. We will develop commercial Series EV-E301 applied this start in March 2014. introduce technical items paper.
This paper tackles a new problem setting: reinforcement learning with pixel-wise rewards (pixelRL) for image processing. After the introduction of deep Q-network, RL has been achieving great success. However, applications processing are still limited. Therefore, we extend to pixelRL various applications. In pixelRL, each pixel an agent, and agent changes value by taking action. We also propose effective method that significantly improves performance considering not only future states own but...
Abstract In this survey, we present a systematic review of 3D hand pose estimation from the perspective efficient annotation and learning. has been an important research area owing to its potential enable various applications, such as video understanding, AR/VR, robotics. However, performance models is tied quality quantity annotated poses. Under status quo, acquiring poses challenging, e.g., due difficulty presence occlusion. To reveal problem, pros cons existing methods classified manual,...
Predicting the users' impressions on a video talk is an important step for recommendation tasks. We propose method to accurately predict multiple impression-related user ratings given talk. Our proposal considers (a) multimodal features including linguistic as well acoustic features, (b) correlations between different (labels), and (c) feature types. In particular, proposed models both label within single Markov random field (MRF), jointly optimizes assignment problem obtain consistent set...
Object affordance is an important concept in hand-object interaction, providing information on action possibilities based human motor capacity and objects' physical property thus benefiting tasks such as anticipation robot imitation learning. However, the definition of existing datasets often: 1) mix up with object functionality; 2) confuse goal-related action; 3) ignore capacity. This paper proposes efficient annotation scheme to address these issues by combining goal-irrelevant actions...
In volume seam carving, i.e., carving for 3D cost volume, an optimal surface can be derived by graph cuts, resulting from sophisticated construction. To date, the graph-cut algorithm is only solution carving. However, it not suitable practical use because incurs a heavy computational load. We propose multipass dynamic programming (DP)-based approach which reduces computation time and memory consumption while maintaining similar image quality as that of cuts. Our DP scheme achieved conducting...
Abstract Hybrid traction systems of lithium‐ion batteries with fuel cells or diesel engines are the one effective measures to reduce emission railcars for nonelectrified lines. Some challenges, including revenue services, hybrid have been recently addressed in Japan. In this paper, evolution and features cell discussed examples challenges. addition, technical trend applications storage is summarized perspectives future systems. Copyright © 2010 Institute Electrical Engineers Published by...
People are fond of taking and sharing photos in their social life, a large part it is face images, especially selfies. A lot researchers interested analyzing attractiveness images. Benefited from deep neural networks (DNNs) training data, have been developing learning models that can evaluate facial photos. However, recent development on DNNs showed they could be easily fooled even when trained dataset. In this paper, we used two approaches to generate adversarial examples high scores but...
Abstract In this work, we present a novel method for simultaneously controlling the head pose and facial expressions of given input image using 3D keypoint-based GAN. Existing methods are not suitable real images, or they generate unnatural results because it is trivial to capture (large changes) (small simultaneously. achieve simultaneous control by introducing keypoints GAN-based synthesis, unlike existing 2D landmark-based approach. As result, our can handle both large variations due...
We present a contrastive learning framework based on in-the-wild hand images tailored for pre-training 3D pose estimators, dubbed HandCLR. Pre-training large-scale achieves promising results in various tasks, but prior methods have not fully utilized the potential of diverse accessible from videos. To facilitate scalable pre-training, we first prepare an extensive pool videos and design our method with learning. Specifically, collected over 2.0M recent human-centric videos, such as 100DOH...
This paper proposes an efficient image retrieval system. When users wish to retrieve images with semantic and spatial constraints (e.g., a horse is located at the center of image, person riding on horse), it difficult for conventional text-based systems such exactly. In contrast, proposed system can consider both information, because based segmentation using fully convolutional networks (FCN). The accept three types as queries: map sketched by user, natural or combination two. distance...
In volume seam carving, carving for three-dimensional (3D) cost volume, an optimal surface can be derived by graph cuts, resulting from sophisticated construction. However, the cuts algorithm is not suitable practical use because it incurs a heavy computational load. We propose multi-pass dynamic programming (DP) based approach that reduces computation time to 60 times faster and memory consumption 10 smaller than those of while maintaining similar image quality as cuts. our DP, suboptimal...
Contextual information such as the co-occurrence of objects and location has played an important role in object detection. We present candidate pruning rescoring methods that leverage contextual can improve state-of-the-art CNN-based detection Fast R-CNN Faster R-CNN. In our method, we formulate reduction a Markov random field optimization problem. employ machine learning technique to reconsider scores windows. experimentally demonstrate improvements R-CNN-based using two datasets. Moreover,...
Japanese comics (called manga) are traditionally created in monochrome format. In recent years, addition to comics, full color a more attractive medium, have appeared. Unfortunately, require manual colorization, which incurs high labor costs. Although automatic colorization methods been recently proposed, most of them designed for illustrations, not comics. Unlike since composed many consecutive images, the painting style must be consistent. To realize consistent we propose here...
Cost-volume filtering is one of the most widely known techniques to solve general multi-label problems, however it problematically inefficient when label space size extremely large. This paper presents a coarse-to-fine strategy cost-volume that handles efficiently and accurately problems with large size. Based upon observation true labels at same image coordinate different scales are highly correlated, we truncate unimportant for by leveraging labeling output lower scales. Experimental...