Search

The fashion clothing and items classification is challenging to incorporate category/sub-category classification and attributes prediction for numerous fashion items into a compact multitask learning infrastructure. The main motive of this research is to improve the fashion items categorization and their attributes prediction from extracted visual features. We proposed a novel fashion sub-categories and attributes prediction (FS C AP) model using deep learning techniques. In this proposed model, YOLO and DeepSORT architectures are used for person detection and tracking, Faster-RCNN architecture is used for sub-categories classification, and Custom-EfficientNet-B3 architecture is designed for attributes prediction. Twenty-four distinct modules are designed to increase the attributes classification accuracy for detected fashion items again each sub-category. The performance of the proposed model is evaluated on a customized fully annotated FashionItem dataset. The experimental results clearly show that the proposed model outperforms the recent baseline methods in fashion sub-categories and attributes prediction. [ABSTRACT FROM AUTHOR]

The optimization of terminal airspace can provide more airport capacity to meet the growing aviation demand. Improving flight procedures is an important prerequisite for optimizing airspace. The existing air-route network visualization cannot fully meet the assessment needs of the terminal flight procedure. In this research, we introduce an analysis tool, VEFP, that provides multiple visualizations based on unlabeled flight trajectory data. The system can help domain experts to evaluate the terminal flight procedure from multiple perspectives, respectively. First, we provide a time series-based statistical information view to help determine the usage status of the flight procedure per unit time. Second, we evaluated the controller's use of space for flight procedures based on the location information in the data. Third, combined with the visualization method after data processing, the visual complexity is reduced and necessary details are displayed. Then, the users can directly observe the actual flight procedure. We evaluate flight procedures in actual use through cases and experiments and then discuss with users about the observations provided by the system. These results confirm that our system can help domain experts evaluate flight procedures. [ABSTRACT FROM AUTHOR]

Modeling large-scale vegetation on terrain is an important and challenging task in computer games, movie production and other digital entertainment applications. In this work, we propose a novel example-based method for rapid generation of vegetation in outdoor natural environments. Its central idea is to learn the vegetation distribution on terrain via deep convolution neural networks. We first use a pre-trained deep neural network to extract rich local information from the terrain pertinent to vegetation distribution. Second, we produce the initial features of the target vegetation distribution based on patch matching and further introduce a network that generates a vegetation density map based on the initial features. Third, during the synthesis stage, we propose a procedural method to generate the vegetation distribution data corresponding to the terrain data. Our research work confirms that the image features extracted by the pre-trained deep neural network could be utilized to explore the connection between vegetation and terrain. We validate our new method over various outdoor scenes, including procedural generated scenes and scenes with manual control on tree patterns. The experimental results demonstrate that our method can rapidly produce new realistic scenes for outdoor natural environments, which relies on the mechanism of learning correlationship between vegetation distribution and terrain data. [ABSTRACT FROM AUTHOR]

Conventional methods to create fluid animation primarily resort to physically based simulation via numerical integration, whose performance is dominantly hindered by large amount of numerical calculation and low efficiency. Alternatively, video-based methods could easily reconstruct fluid surfaces from videos, yet they are not able to realize two-way dynamic interaction with their surrounding environment in a physically correct manner. In this paper, we propose a hybrid method that combines video-based fluid surface reconstruction and popular fluid animation models to compute and re-animate fluid surface. First, the fluid surface's height field corresponding to each video frame is estimated by using the shape-from-shading method. After denoising, hole-filling, and smoothing operations, the height field is utilized to calculate the velocity field, where the shallow water model is adopted. Then we treat the height field and velocity field as real data to drive the simulation. Still, only one layer of surface particles is not capable of driving the smoothed particle hydrodynamics (SPH) system. The surface particles (including 3D position and its velocity) are then employed to guide the spatial sampling of the entire volume underneath. Second, the volume particles corresponding to each video frame are imported into the SPH system to couple with other possible types of particles (used to define interacting objects), whose movement is dictated by the direct forcing method, and fluid particles' geometry information is then corrected by both physical models and real video data. The resulting animation approximates the reconstruction surface from the input video, and new physically based coupling behaviors are also appended. We document our system's detailed implementation and showcase visual performance across a wide range of scenes. [ABSTRACT FROM AUTHOR]

Customizing a desired naturalistic fluid simulation result from video to obtain similar artistic effect is significant in practice. But art-directed customizing is challengeable due to the chaotic nature of the physics contained in it, and this still remains to be a difficult task in spite of rapid advancements of computer graphics during the last two decades. This paper focuses on the problem of physically based fluid re-simulation which is foundational to customize desired naturalistic simulation result from video example. In the previous achievements, conventional algorithms primarily recover 3D geometry of fluid surface or obtain the velocity of fluid particles in video. However, to launch new derivative results, just geometry and velocity are not enough, and physically based driven models are promising. We present a novel method that is capable of efficiently recovering physically driven model from an existing video. We advocate a new approach to calculate the velocity and non-normalized surface geometry under the constraints of the appearance and dynamic behavior of the example fluid in multi-scale framework, and use the calculated physical properties quickly to recover the fluid-driven model. In particular, to calculate the surface geometry more accurately, we introduce a scale factor between normalized geometry and non-normalized geometry to acquire a more accurate result. We propose a novel recovery algorithm, in which the particle densities of lattice Boltzmann method can be recovered more accurately from matching fluid advection geometry with the calculated non-normalized geometry. Fluid re-simulations with different types of density can be achieved, including constrained particle density, auto-advection density, and enhanced particle density. We demonstrate our results in several challenging scenarios and provide qualitative evaluation to our method. Some applications based on our approach are also demonstrated in the implementation. [ABSTRACT FROM AUTHOR]

In this paper, we articulate a novel particle-centric method to simulate the dynamics of brittle fracture for anisotropic materials. The key motivation of this paper is to develop a new hybrid, particle-based simulation that inherits advantages from both powerful finite element methods and popular mesh-free methods, while overcoming certain disadvantages of both types of methods. Our method stems from two novel aspects: (1) a physical model built upon an improved mechanical framework and the adaptive smoothed particle hydrodynamics (SPH), an improved variant of traditional SPH, which can handle complicated anisotropic elastic behaviors with little extra cost; and (2) a hybrid, adaptive particle system that serves for more accurate fracture modeling with richer details. At the physical level, in order to facilitate better control during the formation of fracture and improve its time performance, we develop a physical framework based on contact mechanics and adopt the stress and energy analysis on the anisotropic SPH numerical integration to pinpoint fracture generation and propagation. At the geometric level, in order to reduce time consumption and enhance accuracy in rigid dynamics and fracture generation, we employ hybrid, fully adaptive particles in the vicinity of fracture regions via geometric subdivision. Our novel approach can facilitate the user to control the generation of cracks with low computational cost and retain high-fidelity crack details during animation. Our comprehensive experiments demonstrate the controllability, effectiveness, and accuracy of our method when simulating various brittle fracture patterns for anisotropic materials. [ABSTRACT FROM AUTHOR]