The complexity of online applications is rapidly increasing, bringing more sophisticated performance anomalies in today's cloud datacenter. To fully understand application behaviors, we should obtain both inter-service communication data via RPC-level tracing and intra-service execution traces via application-level tracing to precisely reason about event causality. However, the average time overhead of existing intra-service tracing schemes on the traced applications is generally about 5-10%, possibly reaching 18% in the worst case. To realize practical intra-service tracing in shared and stressed datacenters, one must achieve extreme tracing efficiency with an overhead at the per-mille level.

The past decade has witnessed the huge success of deep learning in well-known artificial intelligence applications such as face recognition, autonomous driving, and large language model like ChatGPT. Recently, the application of deep learning has been extended to a much wider range, with Neural Network-Based Video Coding (NNVC) being one of them. NNVC can be performed at two different levels: embedding neural network-based (NN-based) coding tools into a classical video compression framework or building the entire compression framework upon neural networks. This article elaborates our studies in response to the recent exploration efforts in JVET (Joint Video Experts Team of ITU-T SG 16 WP 3 and ISO/IEC JTC 1/SC29) in the name of NNVC, falling in the former category. Specifically, in this article, we propose two advanced NN-based video coding technologies, i.e., NN-based intra prediction and NN-based in-loop filtering, which have been investigated for several meeting cycles in JVET and then adopted into the reference software, i.e., NNVC. In addition, we further propose a Small Ad-hoc Deep-Learning Library (SADL), which provides integer-based inference capabilities for neural networks to ensure interoperability across different systems. SADL has been adopted as the inference platform of all neural networks in NNVC. Extensive experiments on top of the NNVC have been conducted to evaluate the effectiveness of the proposed techniques. Compared with VTM-11.0_nnvc, the proposed two NN-based coding tools jointly achieve {11.94%, 21.86%, 22.59%}, {9.18%, 19.76%, 20.92%}, and {10.63%, 21.56%, 23.02%} BD-rate reductions on average for {Y, Cb, Cr} under random-access, low-delay, and all-intra configurations, respectively.

Augmented Reality (AR) has been proven beneficial to External Ventricular Drain (EVD) surgery by providing in-situ visual guidance during operations. During this procedure, the key challenge is estimating the spatial relationship between pre-operative images and actual patient anatomy accurately and efficiently. Previous works have revealed conflicts between tracking accuracy, workflow efficiency, and non-invasiveness in tracking pipelines. This research fully utilizes the capabilities of Time of Flight (ToF) depth sensors, including retro-reflective tool tracking and dense surface information, to construct a convenient and accurate EVD guiding pipeline. As previous studies have proven significant depth errors in ToF depth sensors, we first evaluated the feasibility of using ToF sensors in surgical guidance by estimating its accuracy under different conditions and corrected this error in our pipeline. Our results show $ \text{7.580}\pm \text{1.488}\,\text{mm}$7.580±1.488mm depth value errors on human skin under HoloLens 2 depth camera, indicating the significance of depth correction. This error was reduced by over 85% using proposed depth correction method on head phantoms in different materials. The corrected depth information can then be utilized to reconstruct the head surface with sub-millimeter accuracy, validated on a series of 3D-printed models and a sheep head. To demonstrate the effectiveness of the proposed framework, we conducted a case study simulating EVD surgery. Five surgeons were involved in this study, each performing nine k-wire insertions on a head phantom under virtual guidance without tracking for surgical tools. The results revealed $ \text{2.09} \pm \text{1.00}\,\text{mm}$2.09±1.00mm translational and $\text{2.97}\pm \text{1.95}^\circ$2.97±1.95∘ orientational guidance accuracy, demonstrating competitive performance with previous research.

Serverless computing has recently become a significant application paradigm in data centers. However, existing power management methods focus on optimizations at the coarse-grained server level, making them unable to handle the characteristics of these short-lived, dynamic serverless functions. In this context, the unawareness of function-level characteristics by the existing power management systems can severely degrade the energy efficiency of the data centers. To address this challenge, we design a function-level power management system. Instead of relying on server-level schedulers, we propose a novel core-level scheduling policy for serverless functions that can efficiently allocate functions to the most suitable CPU core. Additionally, we propose a power management mechanism for serverless computing that can reduce system power consumption with functions’ QoS guaranteed. Our evaluation shows that our system achieves a maximum power saving of 8.5% and an average power saving of 8% across the majority of loads without incurring any loss in tail latency, as compared to the conventional server-level scheduling system.

The state-of-the-art neural network-based (NN-based) in-loop filters for video coding are built on convolutional neural networks. The Joint Video Experts Team (JVET) activities investigate NN-based in-loop filters for two operation points, the high operation point (HOP) which provides highest possible gains at a high complexity and the low operation point (LOP) which is constrained on a low complexity. This paper focuses on the LOP network. We apply a DCT and reshaping to the inputs and an inverse DCT and inverse reshaping to the outputs of LOP. The spatial resolution inside the network is reduced by a factor of four while the final output still has the same number of pixels. The complexity in MAC/pixel (multiplyaccumulate operations per pixel) is therefore also reduced by a factor of four. This freed-up complexity is instead spent on increasing the number of backbone blocks and channels so the LOP complexity is matched. Our network has a complexity of $16.9 \mathrm{kMAC} /$ pixel and 0.2 M parameters (LOP: 17 kMAC/pixel, 0.05 M parameters). The BD-rate impact compared to the NNVC-7.1 anchor is reported to be −0.48% for RA and −0.17% for AI with the float model, and −0.44% for RA and −0.18% for AI with the integer model.

Human-in-the-loop topic modelling incorporates users' knowledge into the modelling process, enabling them to refine the model iteratively.Recent research has demonstrated the value of user feedback, but there are still issues to consider, such as the difficulty in tracking changes, comparing different models and the lack of evaluation based on real-world examples of use.We developed a novel, interactive human-in-the-loop topic modeling system with a user-friendly interface that enables users compare and record every step they take, and a novel topic words suggestion feature to help users provide feedback that is faithful to the ground truth.Our system also supports not only what traditional topic models can do, i.e., learning the topics from the whole corpus, but also targeted topic modelling, i.e., learning topics for specific aspects of the corpus.In this article, we provide an overview of the system and present the results of a series of user studies designed to assess the value of the system in progressively more realistic applications of topic modelling.

This paper describes a bilateral filter that is being proposed as a coding tool for the Versatile Video Codec (VVC). The filter acts as a loop filter in parallel with the sample-adaptive offset (SAO) filter. Both the proposed filter and SAO act on the same input samples, each filter produces an offset, and these offsets are then added to the input sample to produce an output sample that, after clipping, goes to the next stage. The method has been implemented and tested according to the common test conditions in VVC test model version 5.0. For the all-intra configuration, we report a BD rate figure of -0.4% with an encoder run time increase of 6% and a decoder run time increase of 4%. For the random access configuration, the BD rate figure is -0.5% with an encoder run time increase of 2% and a decoder run time increase of 2%.

Fractional-pel accurate motion is widely used in video coding. For subband coding, fractional-pel accuracy is challenging since it is difficult to handle the complex motion field with temporal transforms. In our previous work, we designed integer accurate motion-adaptive transforms (MAT) which can transform integer accurate motion-connected coefficients. In this paper, we extend the integer MAT to fractional-pel accuracy. The integer MAT allows only one reference coefficient to be the lowband coefficient. In this paper, we design the transform such that it permits multiple references and generates multiple lowband coefficients. In addition, our fractional-pel MAT can incorporate a general interpolation filter into the basis vector, such that the highband coefficient produced by the transform is the same as the prediction error from the interpolation filter. The fractional-pel MAT is always orthonormal. Thus, the energy is preserved by the transform. We compare the proposed fractionalpel MAT, the integer MAT, and the half-pel motion-compensated orthogonal transform (MCOT), while HEVC intra coding is used to encode the temporal subbands. The experimental results show that the proposed fractional-pel MAT outperforms the integer MAT and the half-pel MCOT. The gain achieved by the proposed MAT over the integer MAT can reach up to 1dB in PSNR.

In classic data compression, the optimal transform for energy compaction is the Karhunen-Lòeve transform with the eigenvectors of the covariance matrix. In coding applications, neither the covariance matrix nor the eigenvectors can be easily transmitted to the decoder. In this letter, we introduce a covariance matrix model based on graphs determined by hierarchical block motion in image sequences and use its eigenvector matrix for compression. The covariance matrix model is defined using the graph distance matrix, where the graph is determined by block motion. As the proposed covariance matrix is closely related to the graph, the relation between the covariance matrix and the Laplacian matrix is studied and their eigenvector matrices are discussed. From our assumptions, we show that our covariance model can be viewed as a Gaussian graphical model where the signal is described by the second order statistics and the zeros in the precision matrix indicate missing edges in the graph. To assess the compression performance, we relate the coding gain due to the eigenbasis of the covariance model to that of the Laplacian eigenbasis. The experimental results show that the eigenbasis of our covariance model is advantageous for tree-structured block motion in image sequences.

It is well known that the Karhunen - Loeve Transform (KLT) diagonalizes the covariance matrix and gives the optimal energy compaction. Since the real covariance matrix may not be obtained in video compression, we consider a covariance model that can be constructed without extra cost. In this work, a covariance model based on a graph is considered for temporal transforms of videos. The relation between the covariance matrix and the Laplacian is studied. We obtain an explicit expression of the relation for tree graphs, where the trees are defined by motion information. The proposed graph-based covariance is a good model for motion-compensated image sequences. In terms of energy compaction, our graph-based covariance model has the potential to outperform the classical Laplacian-based signal analysis.

We construct motion-adaptive transforms for image sequences by using the eigenvectors of Laplacian matrices defined on vertex-weighted graphs, where the weights of the vertices are defined by scale factors. The vertex weights determine only the first basis vector of the linear transform uniquely. Therefore, we use these weights to define two Laplacians of vertex-weighted graphs. The eigenvectors of each Laplacian share the first basis vector as defined by the scale factors only. As the first basis vector is common for all considered Laplacians, we refer to it as subspace constraint. The first Laplacian uses the inverse scale factors, whereas the second utilizes the scale factors directly. The scale factors result from the assumption of ideal motion. Hence, the ideal unscaled pixels are equally connected and we are free to form arbitrary graphs, such as complete graphs, ring graphs, or motion-inherited graphs. Experimental results on energy compaction show that the Laplacian which is based on the inverse scale factors outperforms the one which is based on the direct scale factors. Moreover, Laplacians of motion-inherited graphs are superior than that of complete or ring graphs, when assessing the energy compaction of the resulting motion-adaptive transforms.

Motion information in image sequences connects pixels that are highly correlated. In this paper, we consider vertex-weighted graphs that are formed by motion vector information. The vertex weights are defined by scale factors which are introduced to improve the energy compaction of motion-adaptive transforms. Further, we relate the vertex-weighted graph to a subspace constraint of the transform. Finally, we propose a subspace-constrained transform (SCT) that achieves optimal energy compaction for the given constraint. The subspace constraint is derived from the underlying motion information only and requires no additional information. Experimental results on energy compaction confirm that the motion-adaptive SCT outperforms motion-compensated orthogonal transforms while approaching the theoretical performance of the Karhunen Loeve Transform (KLT) along given motion trajectories.

In this paper, we consider motion-adaptive transforms that are based on vertex-weighted graphs. The graphs are constructed by motion vector information and the weights of the vertices are given by scale factors, where the scale factors are used to control the energy compaction of the transform. The vertex-weighted graph defines a one dimensional linear subspace. Thus, our transform basis is subspace constrained. To find a full transform matrix that satisfies our subspace constraint, we rotate the discrete cosine transform (DCT) basis such that the first basis vector matches the subspace constraint. Since rotation is not unique in high dimensions, we choose a simple rotation that only rotates the DCT basis in the plane which is spanned by the first basis vector of the DCT and the subspace constraint. Experimental results on energy compaction show that the motion-adaptive transform based on this rotation is better than the motion-compensated orthogonal transform based on hierarchical decomposition while sharing the same first basis vector.

In this paper, we propose two algorithms to construct motion-adaptive transforms that are based on vertex-weighted graphs. The graphs are constructed by motion vector information. The weights of the vertices are given by scale factors that are used to accommodate proper concentration of energy in transforms. The vertex-weighted graph defines a one dimensional linear subspace. Thus, our transform basis is subspace constrained. We propose two algorithms. The first is based on the Gram-Schmidt orthonormalization of the discrete cosine transform (DCT) basis. The second combines the rotation of the DCT basis and the Gram-Schmidt orthonormalization. We assess both algorithms in terms of energy compaction. Moreover, we compare to prior work on graph-based rotation of the DCT basis and on so-called motion-compensated orthogonal transforms (MCOT). In our experiments, both algorithms outperform MCOT in terms of energy compaction. However, their performance is similar to that of graph-based rotation of the DCT basis.

Well-known standard hybrid coding techniques utilize the concept of motion-compensated predictive coding in a closed-loop. The resulting coding dependencies are a major challenge for packet-based networks like the Internet. On the other hand, subband coding techniques avoid the dependencies of predictive coding and are able to generate video streams that better match packet-based networks. An interesting class for subband coding is the so-called motion-compensated orthogonal transform. It generates orthogonal subband coefficients for arbitrary underlying motion fields. In this paper, a theoretical signal model based on Gaussian distributions is discussed to construct a cost function for efficient rate allocation. Additionally, a rate-distortion efficient video coding scheme is developed that takes advantage of motion-compensated orthogonal transforms. The scheme combines multiple types of motion-compensated orthogonal transforms, variable block sizes, and half-pel accurate motion compensation. The experimental results show that this adaptive scheme outperforms individual motion-compensated orthogonal transforms by up to 2 dB.