A5100447691- Advanced Vision and Imaging
- 3D Shape Modeling and Analysis
- Topic Modeling
- Human Pose and Action Recognition
- Computer Graphics and Visualization Techniques
- Remote Sensing and LiDAR Applications
- Natural Language Processing Techniques
- Multimodal Machine Learning Applications
- Advanced Image and Video Retrieval Techniques
- Advanced Image Processing Techniques
- Advanced Neural Network Applications
- Visual Attention and Saliency Detection
- Anomaly Detection Techniques and Applications
- Video Surveillance and Tracking Methods
- Generative Adversarial Networks and Image Synthesis
- 3D Surveying and Cultural Heritage
- Image Enhancement Techniques
- Image and Video Quality Assessment
- Robotics and Sensor-Based Localization
- Video Coding and Compression Technologies
- Optical measurement and interference techniques
- Software Engineering Research
- Domain Adaptation and Few-Shot Learning
- Image Processing Techniques and Applications
- Advanced Image Fusion Techniques
Peking University
2016-2025
Peking University Shenzhen Hospital
2016-2025
China Academy of Launch Vehicle Technology
2024
University College London
2024
Tianjin University
2022-2024
Peking University First Hospital
2024
Peng Cheng Laboratory
2018-2023
Zhejiang Sci-Tech University
2008-2023
China University of Geosciences
2023
Jianghan University
2023
Relation classification is an important research arena in the field of natural language processing (NLP).In this paper, we present SDP-LSTM, a novel neural network to classify relation two entities sentence.Our architecture leverages shortest dependency path (SDP) between entities; multichannel recurrent networks, with long short term memory (LSTM) units, pick up heterogeneous information along SDP.Our proposed model has several distinct features: (1) The paths retain most relevant (to...
Programming language processing (similar to natural processing) is a hot research topic in the field of software engineering; it has also aroused growing interest artificial intelligence community. However, different from sentence, program contains rich, explicit, and complicated structural information. Hence, traditional NLP models may be inappropriate for programs. In this paper, we propose novel tree-based convolutional neural network (TBCNN) programming processing, which convolution...
Video anomaly detection under weak labels is formulated as a typical multiple-instance learning problem in previous works. In this paper, we provide new perspective, i.e., supervised task noisy labels. such viewpoint, long cleaning away label noise, can directly apply fully action classifiers to weakly detection, and take maximum advantage of these well-developed classifiers. For purpose, devise graph convolutional network correct Based upon feature similarity temporal consistency, our...
In this paper, we propose the TBCNNpair model to recognize entailment and contradiction between two sentences.In our model, a tree-based convolutional neural network (TBCNN) captures sentencelevel semantics; then heuristic matching layers like concatenation, element-wise product/difference combine information in individual sentences.Experimental results show that outperforms existing sentence encoding-based approaches by large margin.
Image inpainting techniques have shown significant improvements by using deep neural networks recently. However, most of them may either fail to reconstruct reasonable structures or restore fine-grained textures. In order solve this problem, in paper, we propose a two-stage model which splits the task into two parts: structure reconstruction and texture generation. first stage, edge-preserved smooth images are employed train reconstructor completes missing inputs. second based on...
Low-light images are not conducive to human observation and computer vision algorithms due their low visibility. To solve this problem, many image enhancement techniques have been proposed. However, existing inevitably introduce color lightness distortion when increasing lower the distortion, we propose a novel method using response characteristics of cameras. First, investigate relationship between two with different exposures obtain an accurate camera model. Then borrow illumination...
Transfer learning is aimed to make use of valuable knowledge in a source domain help model performance target domain.It particularly important neural networks, which are very likely be overfitting.In some fields like image processing, many studies have shown the effectiveness network-based transfer learning.For NLP, however, existing only casually applied learning, and conclusions inconsistent.In this paper, we conduct systematic case provide an illuminating picture on transferability networks NLP. 1
Low-light images are not conducive to human observation and computer vision algorithms due their low visibility. Although many image enhancement techniques have been proposed solve this problem, existing methods inevitably introduce contrast under- over-enhancement. Inspired by visual system, we design a multi-exposure fusion framework for low-light enhancement. Based on the framework, propose dual-exposure algorithm provide an accurate lightness Specifically, first weight matrix using...
Low-light image enhancement algorithms can improve the visual quality of low-light images and support extraction valuable information for some computer vision techniques. However, existing techniques inevitably introduce color lightness distortions when enhancing images. To lower distortions, we propose a novel framework using response characteristics cameras. First, discuss how to determine reasonable camera model its parameters. Then, use illumination estimation estimate exposure ratio...
Using neural networks to generate replies in human-computer dialogue systems is attracting increasing attention over the past few years. However, performance not satisfactory: network tends safe, universally relevant which carry little meaning. In this paper, we propose a content-introducing approach network-based generative systems. We first use pointwise mutual information (PMI) predict noun as keyword, reflecting main gist of reply. then seq2BF, "sequence backward and forward sequences"...
Pose-guided person image generation is to transform a source target pose. This task requires spatial manipulations of data. However, Convolutional Neural Networks are limited by the lack ability spatially inputs. In this paper, we propose differentiable global-flow local-attention framework reassemble inputs at feature level. Specifically, our model first calculates global correlations between sources and targets predict flow fields. Then, flowed local patch pairs extracted from maps...
Fully convolutional neural networks (FCNs) have shown outstanding performance in many computer vision tasks including salient object detection. However, there still remains two issues needed to be addressed deep learning based saliency One is the lack of tremendous amount annotated data train a network. The other robustness for extracting objects images containing complex scenes. In this paper, we present new architecture-PDNet, robust prior-model guided depth-enhanced network RGB-D contrast...
The use of complementary information, namely depth or thermal has shown its benefits to salient object detection (SOD) during recent years. However, the RGB-D RGB-T SOD problems are currently only solved independently, and most them directly extract fuse raw features from backbones. Such methods can be easily restricted by low-quality modality data redundant cross-modal features. In this work, a unified end-to-end framework is designed simultaneously analyze tasks. Specifically, effectively...
Generating portrait images by controlling the motions of existing faces is an important task great consequence to social media industries. For easy use and intuitive control, semantically meaningful fully disentangled parameters should be used as modifications. However, many techniques do not provide such fine-grained controls or indirect editing methods i.e. mimic other individuals. In this paper, a Portrait Image Neural Renderer (PIRenderer) proposed control face with three-dimensional...
In point cloud compression, sufficient contexts are significant for modeling the distribution. However, gathered by previous voxel-based methods decrease when handling sparse clouds. To address this problem, we propose a multiple-contexts deep learning framework called OctAttention employing octree structure, memory-efficient representation Our approach encodes symbol sequences in lossless way gathering information of sibling and ancestor nodes. Expressly, first represent clouds with to...
Saliency detection aims to detect the most attractive objects in images, which has been widely used as a foundation for various multimedia applications. In this paper, we propose novel salient object algorithm RGB-D images using center-dark channel prior. First, generate an initial saliency map based on color and depth of given image. Then, center prior dark Finally, fuse with final map. The proposed is evaluated two public datasets, experimental results show that our method outperforms...
Generic object detection algorithms have proven their excellent performance in recent years. However, on underwater datasets is still less explored. In contrast to generic datasets, images usually color shift and low contrast; sediment would cause blurring images. addition, creatures often appear closely each other due living habits. To address these issues, our work investigates augmentation policies simulate overlapping, occluded blurred objects, we construct a model capable of achieving...
Relation classification is an important research arena in the field of natural language processing (NLP). In this paper, we present SDP-LSTM, a novel neural network to classify relation two entities sentence. Our architecture leverages shortest dependency path (SDP) between entities; multichannel recurrent networks, with long short term memory (LSTM) units, pick up heterogeneous information along SDP. proposed model has several distinct features: (1) The paths retain most relevant (to...
Code generation maps a program description to executable source code in programming language. Existing approaches mainly rely on recurrent neural network (RNN) as the decoder. However, we find that contains significantly more tokens than natural language sentence, and thus it may be inappropriate for RNN capture such long sequence. In this paper, propose grammar-based structural convolutional (CNN) generation. Our model generates by predicting grammar rules of language; design several CNN...
Cloth-changing person re-identification (re-ID) is a new rising research topic that aims at retrieving pedestrians whose clothes are changed. This task quite challenging and has not been fully studied to date. Current works mainly focus on body shape or contour sketch, but they robust enough due view posture variations. The key this exploit cloth-irrelevant cues. paper proposes semantic-guided pixel sampling approach for the cloth-changing re-ID task. We do explicitly define which feature...
In video surveillance, pedestrian retrieval (also called person re-identification) is a critical task. This task aims to retrieve the of interest from non-overlapping cameras. Recently, transformer-based models have achieved significant progress for this However, these still suffer ignoring fine-grained, part-informed information. paper proposes multi-direction and multi-scale Pyramid in Transformer (PiT) solve problem. architecture, each image split into many patches. Then, patches are fed...
With the prevalence of thermal cameras, RGB-T multi-modal data have become more available for salient object detection (SOD) in complex scenes. Most SOD works first individually extract RGB and features from two separate encoders directly integrate them, which pay less attention to issue defective modalities. However, such an indiscriminate feature extraction strategy may produce contaminated thus lead poor performance. To address this issue, we propose a novel CCFENet perspective perform...
Point clouds upsampling is a challenging issue to gener-ate dense and uniform point from the given sparse input. Most existing methods either take end-to-end su-pervised learning based manner, where large amounts of pairs input ground-truth are exploited as supervision information; or treat up-scaling different scale factors independent tasks, have build multiple networks handle with varying factors. In this paper, we propose novel approach that achieves self-supervised...
Sixth-generation (6G) networks are evolving toward new features and order-of-magnitude enhancement of systematic performance metrics compared to the current 5G. In particular, 6G expected achieve extreme connectivity with Tbps-scale data rate, Kbps/Hz-scale spectral efficiency, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu\mathrm{s}$</tex> , -scale latency. To this end, an original three-layer network architecture is designed realize...