The unpredictably increasing demand of volumetric video streaming for capturing immersive objects has enabled today's multimedia research community to focus on video streaming in six-degree-of-freedom (6DoF). However, there is no one-size-fits-all solution for an end-to-end (from capturing to rendering) volumetric 6DoF video streaming over HTTP. Therefore, this paper presents an architectural approach for an end-to-end 6DoF video streaming, leveraging the de-facto HTTP-based streaming protocol over Internet (a.k.a Dynamic Adaptive Streaming Over HTTP; DASH). The envision architecture depicts and realizes an end-to-end video streaming architecture from 6DoF media capturing to its rendering at the client-side. The proposed architecture complies with existing Internet architecture and consists of three essential components such as: 1) Video Provisioning, 2) Delivery, and 3) Consumption. The architecture exhibits future optimizations to the researchers and practitioners of volumetric video delivery.
Recently, the research community of HTTP-based video delivery (known as HTTP Adaptive streaming; HAS) has proposed many adaptive bitrate (ABR) techniques for dynamically selecting suitable bitrates and adapting to the available network conditions. However, it is imperative to understand the working principle of these ABR techniques. In this paper, we investigated client-driven heuristic-based ABR techniques that use buffer level (a.k.a buffer-based) or estimated throughput (a.k.a throughput-based) to select bitrate for video-on-demand delivery services. To do so, we used the open-source dash.js reference player which implements these techniques. For performance evaluation, the experiments are conducted using various video sessions and under different network conditions, in which we have collected buffer level, selected bitrate level, and latency for each video session. Our result analysis reveals that the throughput-based technique outperforms the buffer-based technique when the target is to achieve the highest video bitrate with fewer rebuffering events.
Nowadays, users and large corporations are looking for strong deep learning models that requires large amount of data; however, they are reluctant to share their private data. Federated learning (FL) enables decentralized training on decentralized data. Without the need to share the data, users can train machine learning models locally, and share the model instead. The main issue with the server-based federated learning is the centralization of governance, the server has a full control on the federated learning process. To solve this issue, a blockchain-based federated learning system is proposed, the blockchain will enable trust in a fully decentralized federated learning environment.
Android is the most widespread smartphone operating system. Its popularity attracted attackers to develop all sorts of malicious applications. On the defense side, much research has been done toward identifying Android applications type and state based on their system-level behavior. This paper describes a Machine Learning technique to detect whether an application is currently running in the foreground or not. The technique is aimed at boosting the accuracy of behavioral malware detection, by identifying apps performing suspect I/O activities while running in the background. We report that a lightly trained ML model focusing on a limited number of system calls over several applications is able to identify the running application along with its state with high confidence.
This paper presents a novel hybrid CMOS-Memristor-based in-memory processing architecture of the Random Spray Retinex algorithm for Imaging applications. The presented approach supports parallel computing and provides efficiency gains in area and power compared to traditional implementation. It uses the analog computations in the memristor crossbar and exploits the same physical elements for both processing and storage. As a result, it substantially reduces the computational complexity resulting from the data-intensive algorithm. Random Spray Retinex is presented as a case study. The experiments suggest that the parallel memristor-based in-memory computation is highly efficient in achieving a high resolution while greatly accelerating the RSR algorithm. The proposed approach is considered a milestone towards low-complexity and real-time Hardware architecture and design of the retinex algorithm.
More content providers use HTTP Adaptive Streaming have started delivering high quality streams with advanced end-to-end encryption mechanisms. This increase in HAS encrypted traffic, creates a significant challenge for network providers in understanding what is happening on their infrastructures which limits their ability to manage network infrastructures properly. Accordingly, the network providers could not take appropriate decisions for better optimizations, resulting in significant revenue lost. Inferring the quality of experience (QoE) of HAS-based streaming video services which mostly rely on packet inspections, showing low performance in inference accuracy. To address this issue, we develop a machine learning powered system that infers QoE factors such as startup delay, rebuffering and selected quality. Our solution uses Deep Self Organizing Map (DSOM) and Multi Layer Perceptron Backpropagation (MLPB), allowing efficient accuracy with low error in inferring QoE factors over several public video datasets, compared to some state-of-the-art approaches.
Retinal vessel segmentation is a critical biomedical image processing task, it requires high accuracy. There are two types of vessels, thick vessels and thin vessels. The challenge lays in the segmentation of the thin vessel branches. Another challenge in the class imbalance between vessel pixels and non-vessel pixels. As the non-vessel pixels outnumber the thin lines of pixels of the vessels [1]. To reduce the effect of the class imbalance, a proposed segmentation model will be the U-net architecture trained to multiple datasets. The datasets are DRIVE, STARE, HRF, and CHASE-DB. The datasets were initially resized to DRIVE dataset dimensions. Then, the model was trained with all datasets. Lastly it was tested on all datasets together and individually. The F1-score has increases by 0.84%, 1.94%, and 0.27%, for STARE, HRF and CHASE-DB. The sensitivity of the model showed improvement in 2.46%, 7.1%, and 3% for STARE, HRF and CHASE-DB.
Genome imputation is an approach that builds on the statistical grounds of hidden Markov models to facilitate Genome-wide association studies (GWAS) at a lower cost. It uses the haplotype patterns in a reference panel to estimate the missing SNPs in a study target. In this study, we show that the use of a local UAE panel provides higher imputation results imputing UAE samples than existing international reference panels. The usage of the UAE reference panel resulted in concordance value of 98.6% as opposed to 97.1% for the 1000 genome and 96.4% for the combined panel in the case of chromosome 17, and has shown a higher number of correctly imputed variants using the UAE reference panel across all chromosomes for minor allele frequency below 0.5. Achieved results using IMPUTE2, 120 WGS, 1000 genotype array data sets provided by Khalifa University.
This work experimentally investigates the mechanical behavior of a novel modified auxetic honeycomb structure. In auxetic honeycomb structures, the internal members of the cell shaped as 'V' yields undesired interactions causing distortion and instability of the structure under compression. This issue is addressed by flattening the internal members through the addition of an extra member. The new design results in a significant improvement in mechanical properties through the introduction of a secondary linear elastic region. In this work, we experimentally investigate the elastic behavior of the new design in attempt to expand the investigation to include analytical modeling and FEM validation. In addition, we introduce a Fusion Deposition Modeling (FDM) 3D printing process that yields isotropic structures.
Bucky papers with thickness of more 400 ?m are very important for many applications such as Radar Absorbing Materials (RAM), Electromagnetic interference shielding (EMI), supercapacitors and lithium-ion battery (LIB). CNT Buckypaper is a thin sheet made from an aggregate of carbon nanotubes, literature studies show the thickness of single bucky paper produced by film casting method is in the range of 40-100 ?m1. This study shares the preliminary work carried as an attempt to produce super-thick bucky papers by film casting process.
