A5100755394- Multimodal Machine Learning Applications
- Advanced Image and Video Retrieval Techniques
- Image and Object Detection Techniques
- Optical measurement and interference techniques
- Domain Adaptation and Few-Shot Learning
- Advanced Vision and Imaging
- Anomaly Detection Techniques and Applications
- Video Analysis and Summarization
- Advanced Graph Neural Networks
- Machine Learning and Algorithms
- Image Processing Techniques and Applications
- Generative Adversarial Networks and Image Synthesis
- Image Processing and 3D Reconstruction
- 3D Surveying and Cultural Heritage
- Advanced Neural Network Applications
- Human Pose and Action Recognition
- Model Reduction and Neural Networks
- Advanced Measurement and Metrology Techniques
- Adversarial Robustness in Machine Learning
- Digital Marketing and Social Media
- Structural Health Monitoring Techniques
- Fault Detection and Control Systems
- Fluid Dynamics and Turbulent Flows
- Graph Theory and Algorithms
- Reinforcement Learning in Robotics
Chinese Academy of Medical Sciences & Peking Union Medical College
2024
Alibaba Group (China)
2020-2023
University of Huddersfield
2022
Hebei University of Technology
2019-2021
Alibaba Group (United States)
2021
Hainan University
2021
Shantou University
2021
Nanjing University
2012-2020
Shanghai University
2019
The University of Adelaide
2011-2014
The ability to generate good model hypotheses is instrumental accurate and robust geometric fitting. We present a novel dynamic hypothesis generation algorithm for fitting of multiple structures. Underpinning our method fast guided sampling scheme enabled by analysing correlation preferences induced data residuals. Our progressively accumulates evidence in the search space, uses information dynamically (1) identify outliers, (2) filter unpromising hypotheses, (3) bias active discovery...
Random hypothesis generation is central to robust geometric model fitting in computer vision. The predominant technique randomly sample minimal subsets of the data, and hypothesize models from selected subsets. While taking increases chance successively “hitting” inliers a sample, hypotheses fitted on may be severely biased due influence measurement noise, even if contain purely inliers. In this paper we propose Cluster Models, used simulate coupled spin systems, conduct using larger than...
In this paper, we propose to investigate the problem of out-of-domain visio-linguistic pretraining, where pretraining data distribution differs from that downstream on which pretrained model will be fine-tuned. Existing methods for are purely likelihood-based, leading spurious correlations and hurt generalization ability when transferred tasks. By correlation, mean conditional probability one token (object or word) given another can high (due dataset biases) without robust (causal)...
Camera and projector are the key components of structured light three-dimensional (3-D) measurements, Digital Light Processing (DLP) has been widely used for projecting digital patterns measurement. The projectors can be modeled as inverse procedures camera imaging, its high-accuracy calibration is still a remaining challenge. Therefore, this paper proposes novel method to improve accuracy DLP projector. By fixing position board, essentially eliminates perspective transformation error...
In the field of three-dimensional (3-D) metrology based on fringe projection profilometry (FPP), accurate camera calibration is an essential task and a primary requirement. order to improve accuracy calibration, board or target needs be manufactured with high accuracy, marker points in image require positioned accuracy. This paper presents improved method by simultaneously optimizing parameters geometry. Specifically, set regularly distributed markers rich coded concentric ring pattern first...
In e-commerce, a growing number of user-generated videos are used for product promotion. How to generate video descriptions that narrate the user-preferred characteristics depicted in is vital successful promoting. Traditional captioning methods, which focus on routinely describing what exists and happens video, not amenable product-oriented captioning. To address this problem, we propose captioner framework, abbreviated as Poet. Poet firstly represents spatial-temporal graphs. Then, based...
In e-commerce, consumer-generated videos, which in general deliver consumers' individual preferences for the different aspects of certain products, are massive volume. To recommend these videos to potential consumers more effectively, diverse and catchy video titles critical. However, seldom accompany appropriate titles. bridge this gap, we integrate comprehensive sources information, including content narrative comment sentences supplied by consumers, product attributes, an end-to-end...
We present a method for learning max-weight matching predictors in bipartite graphs. The consists of performing maximum posteriori estimation exponential families with sufficient statistics that encode permutations and data features. Although inference is general hard, we show one very relevant application - web page ranking exact efficient. For model instances, an appropriate sampler readily available. Contrary to existing max-margin models, our approach statistically consistent and,...
Random hypothesis generation is central to robust geometric model fitting in computer vision. The predominant technique randomly sample minimal or elemental subsets of the data, and hypothesize from selected subsets. While taking increases chance simultaneously "hitting" inliers a sample, it amplifies noise underlying model, hypotheses fitted on may be severely biased even if they contain purely inliers. In this paper we propose use Cluster Models, used simulate coupled spin systems, conduct...
In recommender systems, click behaviors play a fundamental role in mining users’ interests and training models (clicked items as positive samples). Such signals are implicit feedback arguably less representative of inherent interests. Most existing works denoise by introducing external signals, such gaze, dwell time, “like” behaviors. However, explicit is not always routinely available, or might be problematic to collect on large scale. this paper, we identify that an interaction’s related...
In recent years, online short videos have become more popular, especially as an advertising intermediary. To better understand their effects advertisements, it is essential to analyze the causal relations of on consumer behaviors. Our study based fine-grained behavior data from a world-leading e-commerce platform, i.e., Taobao.com. We first decompose total into informative and persuasive following common practice in economic literature. Moreover, we extract subjectivity scores through...
A quantile binary classifier uses the rule: Classify x as +1 if P(Y = 1|X x) >= t, and -1 otherwise, for a fixed parameter t {[0, 1]. It has been shown that Support Vector Machines (SVMs) in limit are classifiers with 1/2 . In this paper, we show by using asymmetric cost of misclassification SVMs can be appropriately extended to recover, limit, any t. We then present principled algorithm solve SVM all values simultaneously. This two implications: First, one recover entire conditional...
In this paper, we propose a neural network based approach to approximate numerical solutions of partial differential equations (PDEs). Concretely, understand the solution PDEs as finding reasonable mapping from with initial and boundary conditions space solutions. Then, by best fitting data PDEs, where are cooperated loss function. The main contributions proposed method twofold. First, adopt quasi-Newton algorithm minimize instead stochastic gradient method, which is benefit relatively low...