We show that relation modeling between visual elements matters in cropping view recommendation. Cropping recommendation addresses the problem of image recomposition conditioned on composition quality and ranking views (cropped sub-regions). This task is challenging because difference subtle when a element reserved or removed. Existing methods represent by extracting region-based convolutional features inside outside boundaries, without probing fundamental question: why some are interest...

Crowd counting based on density maps is generally regarded as a regression task.Deep learning used to learn the mapping between image content and crowd distribution. Although great success has been achieved, some pedestrians far away from camera are difficult be detected. And number of hard examples often larger. Existing methods with simple Euclidean distance algorithm indiscriminately optimize easy so that densities usually incorrectly predicted lower or even zero, which results in large...

The network intrusion detection systems which based on anomaly techniques plays an important role in protection and from harmful attacks.With increasing attacks the new security challenges, lower accuracy of method cluster analysis traffic is a big question, In this paper, we proposed hybrid by combining Particle Swarm Optimization(PSO) K-Means clustering algorithms improving accuracy.We first preprocess features data traffic, extract characteristics various categories attack, then use...

Many scientific applications have started using deep learning methods for their classification or regression problems. However, data-intensive applications, I/O performance can be the major bottleneck. In order to effectively solve important real-world problems on High-Performance Computing (HPC) systems, it is essential address poor issue in large-scale neural network training. this paper, we propose an asynchronous strategy that generally applied applications. Our employs -dedicated thread...

Neural networks are powerful solutions to many scientific applications; however, they usually require long model training time due large data sets or size. Research has been focused on developing numerical optimization algorithms and parallel processing reduce the time. In this work, we propose a multi-resolution strategy that can by with reduced-resolution samples at beginning later switching original resolution samples. This is motivated observation coarser versions of applications be...

Current 3D stylization methods often assume static scenes, which violates the dynamic nature of our real world. To address this limitation, we present S-DyRF, a reference-based spatio-temporal method for neural radiance fields. However, stylizing scenes is inherently challenging due to limited availability stylized reference images along temporal axis. Our key insight lies in introducing additional cues besides provided reference. end, generate pseudo-references from given These facilitate...

Introduction Reconstructing low-level particle tracks in neutrino physics can address some of the most fundamental questions about universe. However, processing petabytes raw data using deep learning techniques poses a challenging problem field High Energy Physics (HEP). In Exa.TrkX Project, an illustrative HEP application, preprocessed simulation is fed into state-of-art Graph Neural Network (GNN) model, accelerated by GPUs. limited GPU memory often leads to Out-of-Memory (OOM) exceptions...

Traversability estimation is the foundation of path planning for a general navigation system. However, complex and dynamic environments pose challenges latest methods using self-supervised learning (SSL) technique. Firstly, existing SSL-based generate sparse annotations lacking detailed boundary information. Secondly, their strategies focus on hard samples rapid adaptation, leading to forgetting biased predictions. In this work, we propose IMOST, continual traversability framework composed...

Mini-batch shuffling is important for the deep learning training process. Most people use random method, which aims to produce a permutation of dataset in every epoch. In this study, we explore mini-batch multi-class imbalanced data classification by investigating several strategies. We find that different order input can significantly affect results models. The show our proposed strategies improve accuracy around 2%, demonstrating higher diversity and lower imbalance ratio each lead better results.

Learning accurate object detectors often requires large-scale training data with precise bounding boxes. However, labeling such is expensive and time-consuming. As the crowd-sourcing process ambiguities of objects may raise noisy box annotations, will suffer from degenerated data. In this work, we aim to address challenge learning robust inaccurate Inspired by fact that localization precision suffers significantly boxes while classification accuracy less affected, propose leveraging as a...

The Segment Anything Model (SAM) has demonstrated its effectiveness in segmenting any part of 2D RGB images. However, SAM exhibits a stronger emphasis on texture information while paying less attention to geometry when To address this limitation, we propose the Any RGBD (SAD) model, which is specifically designed extract directly from Inspired by natural ability humans identify objects through visualization depth maps, SAD utilizes segment rendered map, thus providing cues with enhanced and...

Class-agnostic motion prediction methods aim to comprehend within open-world scenarios, holding significance for autonomous driving systems. However, training a high-performance model in fully-supervised manner always requires substantial amounts of manually annotated data, which can be both expensive and time-consuming obtain. To address this challenge, our study explores the potential semi-supervised learning (SSL) class-agnostic prediction. Our SSL framework adopts consistency-based...

As the country pays more and attention to garbage sorting recycling, people should raise their awareness of environmental protection. This paper proposes an intelligent APP based on Android Studio. The uses Client-Server mode work with computers, SQLite database technology JAVA language write programs. And it realizes a variety applications, such as popularize knowledge classification liking community experience classification, monitoring recycling by users calculating points exchange goods,...

Human activity recognition (HAR) has a wide application in daily life. With wearable sensors, people’s can be monitored, recorded and analysed. However, most existing methods did not make full use of human data their features. In this paper, new method based on feature selection clustering algorithm is proposed. We established two-layer classification models, respectively, the ankle, chest, wrist mixed accelerometer data. K-means was first used to obtain broad activities then we conducted...

In High Energy Physics (HEP), experimentalists generate large volumes of data that, when analyzed, helps us better understand the fundamental particles and their interactions. This is often captured in many files small size, creating a management challenge for scientists. order to facilitate management, transfer, analysis on scale platforms, it advantageous aggregate further into smaller number larger files. However, this translation process can consume significant time resources, if...