A5073315875- Advanced Neural Network Applications
- Domain Adaptation and Few-Shot Learning
- Image Enhancement Techniques
- Advanced Image and Video Retrieval Techniques
- Advanced Image Fusion Techniques
- Image and Signal Denoising Methods
- Multimodal Machine Learning Applications
- Music and Audio Processing
- Speech and Audio Processing
- Advanced Vision and Imaging
- Adversarial Robustness in Machine Learning
- Robotics and Sensor-Based Localization
- COVID-19 diagnosis using AI
- Robotic Locomotion and Control
- Image Processing Techniques and Applications
- Advanced Image Processing Techniques
- Hand Gesture Recognition Systems
- Visual Attention and Saliency Detection
- Human Pose and Action Recognition
- Anomaly Detection Techniques and Applications
- stochastic dynamics and bifurcation
- Robot Manipulation and Learning
- Water Systems and Optimization
- Recommender Systems and Techniques
- Epigenetics and DNA Methylation
Chinese University of Hong Kong, Shenzhen
2017-2025
Institute of Art
2021-2023
Chinese University of Hong Kong
2020
The width of a neural network matters since increasing the will necessarily increase model capacity. However, performance does not improve linearly with and soon gets saturated. In this case, we argue that number networks (ensemble) can achieve better accuracy-efficiency trade-offs than purely width. To prove it, one large is divided into several small ones regarding its parameters regularization components. Each these has fraction original one's parameters. We then train together make them...
Audio-Visual scene understanding is a challenging problem due to the unstructured spatial-temporal relations that exist in audio signals and spatial layouts of different objects visual images. Recently, many studies have focused on abstracting features from convolutional neural networks while learning explicit semantically relevant frames sound images has been overlooked. To this end, we present an end-to-end framework, namely attentional graph network (AGCN), for structure-aware...
Accurate perception of the surrounding scene is helpful for robots to make reasonable judgments and behaviours. Therefore, developing effective representation recognition methods are significant importance in robotics. Currently, a large body research focuses on novel auxiliary features networks improve indoor ability. However, few them focus directly constructing object relations recognition. In this paper, we analyze weaknesses current propose an Object-to-Scene (OTS) method, which...
Natural image matting aims to precisely separate foreground objects from backgrounds using alpha mattes. Fully automatic natural without external annotations is challenging. Well-performed methods usually require accurate labor-intensive handcrafted trimap as an extra input while the performance of generation method, e.g., erosion/dilation manipulation on segmentation, fluctuates with segmentation quality. Therefore, we argue that how produce a high-quality coarse major issue in matting. In...
The ability to maintain balance and stability is a major challenge for quadruped robots in locomotion. For example, when walking the unstructured environment, robot not only needs adjust its gait according terrain information, but also tune body posture based on real-time acquired from inertial measurement unit (IMU) stability. In this paper, we present method that focuses design of correction controller IMU feedback adaptive slope robot. To end, integrated into foot trajectory generator...
Scene recognition is a fundamental task in robotic perception. For human beings, scene reasonable because they have abundant object knowledge of the real world. The idea transferring prior from humans to significant but still less exploited. In this paper, we propose utilize meaningful representations for indoor representation. First, an improved model (IOM) as baseline that enriches by introducing parsing algorithm pretrained on ADE20K dataset with rich categories related scene. To analyze...
Humans have a remarkable ability to learn continuously from th e environment and inner experience. One of the grand goals robots is build an artificial "lifelong learning" agent that can shape cultivated understanding world current scene previous knowledge via autonomous lifelong development. It challenging for robot learning process retain earlier when encounter new tasks or information. Recent advances in computer vision deep -learning methods been impressive due large-scale data sets,...
Accurate turning is not an easy task for the quadruped robot and previous work mainly addresses problem with complex control strategy or systems. This paper proposes a design method accurate motion of 12 degrees freedom (DOF). Unlike others, we have achieved precise relatively simple geometric approach. Once parameters desired movement are set, such as radius angles, can realize corresponding circle locomotion. A architecture including analysis, 3D foot trajectory inverse kinematics...
The hand gesture recognition plays an important role in the human-computer interaction (HCI) field. Previous works mainly focused on researching this task shorter distance, which can only be applied a limited scenario due to distance constrains. In paper, we propose method, is named Joint Single Shot Multibox Detector (JSSD) network, solve hard long-distance task. method based framework of SSD. Our model employed two SSDs for object detection. One head-shoulder region detection and other...
Environmental sound classification (ESC) is a challenging problem due to the unstructured spatial-temporal relations that exist in signals. Recently, many studies have focused on abstracting features from convolutional neural networks while learning of semantically relevant frames signals has been overlooked. To this end, we present an end-to-end framework, namely feature pyramid attention network (FPAM), focusing for ESC. We first extract maps preprocessed spectrogram waveform by backbone...
This paper presents the gait design and comparison study of a quadruped robot. The main contributions this are: (i) Explore modularized system an 8 degree-of-freedom robot with power management system. (ii) Propose method based on drive function. (iii) another foot trajectory. (iv) simulation walk trot gaits above-mentioned methods are validated in Webots. (v) successfully implemented real Experimental results show that stable locomotion can be achieved improved elliptical
Hand gesture recognition plays an essential role in the human-robot interaction (HRI) field. Most previous research only studies hand a short distance, which cannot be applied for with mobile robots like unmanned aerial vehicles (UAVs) at longer and safer distance. Therefore, we investigate challenging long-range problem between humans UAVs. To this end, propose novel attention-based single shot multibox detector (SSD) model that incorporates both spatial channel attention recognition. We...
Personal handwriting style fonts generation is a diverting but time-consuming task due to the large size of Chinese character set. In addition, unlike standard printed fonts, hand-writing are more complicated stroke and glyph feature. this paper, an improved network architecture proposed for learning personal based on small The composed three sub-networks: 1) classification identifying general target fonts; 2) generating transferring identified 3) discriminating differentiating generated...
Accurate perception of the surrounding scene is helpful for robots to make reasonable judgments and behaviours. Therefore, developing effective representation recognition methods are significant importance in robotics. Currently, a large body research focuses on novel auxiliary features networks improve indoor ability. However, few them focus directly constructing object relations recognition. In this paper, we analyze weaknesses current propose an Object-to-Scene (OTS) method, which...
Image classification plays a pivotal role across diverse applications, yet challenges persist when models are deployed in real-world scenarios. Notably, these falter detecting unfamiliar classes that were not incorporated during classifier training, formidable hurdle for safe and effective model deployment, commonly known as out-of-distribution (OOD) detection. While existing techniques, like max logits, aim to leverage logits OOD identification, they often disregard the intricate interclass...
Abstract Audio‐visual scene classification (AVSC) poses a formidable challenge owing to the intricate spatial‐temporal relationships exhibited by audio‐visual signals, coupled with complex spatial patterns of objects and textures found in visual images. The focus recent studies has predominantly revolved around extracting features from diverse neural network structures, inadvertently neglecting acquisition semantically meaningful regions crucial components within data. authors present...
Unbiased scene graph generation (USGG) is a challenging task that requires predicting diverse and heavily imbalanced predicates between objects in an image. To address this, we propose novel framework peer learning uses predicate sampling consensus voting (PSCV) to encourage multiple peers learn from each other. Predicate divides the classes into sub-distributions based on frequency, assigns different handle sub-distribution or combinations of them. Consensus ensembles peers' complementary...
In this paper, we extend the blind-spot based self-supervised denoising by using affinity learning to remove noise from affected pixels. Inspired inpainting, introduce a novel Mask Guided Residual Convolution (MGRConv) learn neighboring image pixel map that gradually removes and refines process. We show mask convolution plays an important role in since it is theoretically aligned with $\mathcal{J} - invariance$, which frameworks are built upon. The theoretical analysis further shows...
With the rapid advancements in autonomous driving and robot navigation, there is a growing demand for lifelong learning (LL) models capable of estimating metric (absolute) depth. LL approaches potentially offer significant cost savings terms model training, data storage, collection. However, quality RGB images depth maps sensor-dependent, real world exhibit domain-specific characteristics, leading to variations ranges. These challenges limit existing methods scenarios with small domain gaps...
This report summarizes IROS 2019-Lifelong Robotic Vision Competition (Lifelong Object Recognition Challenge) with methods and results from the top $8$ finalists (out of over~$150$ teams). The competition dataset (L)ifel(O)ng (R)obotic V(IS)ion (OpenLORIS) - (OpenLORIS-object) is designed for driving lifelong/continual learning research application in robotic vision domain, everyday objects home, office, campus, mall scenarios. explicitly quantifies variants illumination, object occlusion,...
Scene graph generation (SGG) has gained tremendous progress in recent years. However, its underlying long-tailed distribution of predicate classes is a challenging problem. For extremely unbalanced distributions, existing approaches usually construct complicated context encoders to extract the intrinsic relevance scene predicates and complex networks improve learning ability network models for highly imbalanced distributions. To address unbiased SGG problem, we introduce simple yet effective...