Yiannis Andreopoulos

Shakarim Soltanayev, Odysseas Zisimopoulos, Mohammad Ashraful Anam et al.

Video-on-demand streaming has benefitted from \textit{content-adaptive encoding} (CAE), i.e., adaptation of resolution and/or quantization parameters for each scene based on convex hull optimization. However, CAE is very challenging to develop and deploy for interactive game streaming (IGS). Commercial IGS services impose ultra-low latency encoding with no lookahead or buffering, and have extremely tight compute constraints for any CAE algorithm execution. We propose the first CAE approach for resolution adaptation in IGS based on compact encoding metadata from past frames. Specifically, we train a convolutional neural network (CNN) to infer the best resolution from the options available for the upcoming scene based on a running window of aggregated coding block statistics from the current scene. By deploying the trained CNN within a practical IGS setup based on HEVC encoding, our proposal: (i) improves over the default fixed-resolution ladder of HEVC by 2.3 Bjøntegaard Delta-VMAF points; (ii) infers using 1ms of a single CPU core per scene, thereby having no latency overhead.

Alhabib Abbas, Yiannis Andreopoulos

We propose a novel mixture-of-experts class to optimize computer vision models in accordance with data transfer limitations at test time. Our approach postulates that the minimum acceptable amount of data allowing for highly-accurate results can vary for different input space partitions. Therefore, we consider mixtures where experts require different amounts of data, and train a sparse gating function to divide the input space for each expert. By appropriate hyperparameter selection, our approach is able to bias mixtures of experts towards selecting specific experts over others. In this way, we show that the data transfer optimization between visual sensing and processing can be solved as a convex optimization problem.To demonstrate the relation between data availability and performance, we evaluate biased mixtures on a range of mainstream computer vision problems, namely: (i) single shot detection, (ii) image super resolution, and (iii) realtime video action classification. For all cases, and when experts constitute modified baselines to meet different limits on allowed data utility, biased mixtures significantly outperform previous work optimized to meet the same constraints on available data.

Yin Bi, Aaron Chadha, Alhabib Abbas et al.

Neuromorphic vision sensing (NVS)\ devices represent visual information as sequences of asynchronous discrete events (a.k.a., "spikes") in response to changes in scene reflectance. Unlike conventional active pixel sensing (APS), NVS allows for significantly higher event sampling rates at substantially increased energy efficiency and robustness to illumination changes. However, feature representation for NVS is far behind its APS-based counterparts, resulting in lower performance in high-level computer vision tasks. To fully utilize its sparse and asynchronous nature, we propose a compact graph representation for NVS, which allows for end-to-end learning with graph convolution neural networks. We couple this with a novel end-to-end feature learning framework that accommodates both appearance-based and motion-based tasks. The core of our framework comprises a spatial feature learning module, which utilizes residual-graph convolutional neural networks (RG-CNN), for end-to-end learning of appearance-based features directly from graphs. We extend this with our proposed Graph2Grid block and temporal feature learning module for efficiently modelling temporal dependencies over multiple graphs and a long temporal extent. We show how our framework can be configured for object classification, action recognition and action similarity labeling. Importantly, our approach preserves the spatial and temporal coherence of spike events, while requiring less computation and memory. The experimental validation shows that our proposed framework outperforms all recent methods on standard datasets. Finally, to address the absence of large real-world NVS datasets for complex recognition tasks, we introduce, evaluate and make available the American Sign Language letters (ASL-DVS), as well as human action dataset (UCF101-DVS, HMDB51-DVS and ASLAN-DVS).

Abdullahi Mohammad, Christos Masouros, Yiannis Andreopoulos

This paper introduces a framework for systematic complexity scaling of deep neural network(DNN) based MIMO detectors. The model uses a fraction of the DNN inputs by scaling their values through weights that follow monotonically non-increasing functions. This allows for weight scaling across and within the different DNN layers in order to achieve accuracy-vs.-complexity scalability during inference. In order to further improve the performance of our proposal, we introduce a sparsity-inducing regularization constraint in conjunction with trainable weight-scaling functions. In this way, the network learns to balance detection accuracy versus complexity while also increasing robustness to changes in the activation patterns, leading to further improvement in the detection accuracy and BER performance at the same inference complexity. Numerical results show that our approach is 10-foldand 100-fold less complex than classical approaches based on semi-definite relaxation and ML detection, respectively.

Yin Bi, Aaron Chadha, Alhabib Abbas et al.

Neuromorphic vision sensing (NVS)\ devices represent visual information as sequences of asynchronous discrete events (a.k.a., ``spikes'') in response to changes in scene reflectance. Unlike conventional active pixel sensing (APS), NVS allows for significantly higher event sampling rates at substantially increased energy efficiency and robustness to illumination changes. However, object classification with NVS streams cannot leverage on state-of-the-art convolutional neural networks (CNNs), since NVS does not produce frame representations. To circumvent this mismatch between sensing and processing with CNNs, we propose a compact graph representation for NVS. We couple this with novel residual graph CNN architectures and show that, when trained on spatio-temporal NVS data for object classification, such residual graph CNNs preserve the spatial and temporal coherence of spike events, while requiring less computation and memory. Finally, to address the absence of large real-world NVS datasets for complex recognition tasks, we present and make available a 100k dataset of NVS recordings of the American sign language letters, acquired with an iniLabs DAVIS240c device under real-world conditions.

Eirina Bourtsoulatze, Aaron Chadha, Ilya Fadeev et al.

Several groups are currently investigating how deep learning may advance the state-of-the-art in image and video coding. An open question is how to make deep neural networks work in conjunction with existing (and upcoming) video codecs, such as MPEG AVC, HEVC, VVC, Google VP9 and AOM AV1, as well as existing container and transport formats, without imposing any changes at the client side. Such compatibility is a crucial aspect when it comes to practical deployment, especially due to the fact that the video content industry and hardware manufacturers are expected to remain committed to these standards for the foreseeable future. We propose to use deep neural networks as precoders for current and future video codecs and adaptive video streaming systems. In our current design, the core precoding component comprises a cascaded structure of downscaling neural networks that operates during video encoding, prior to transmission. This is coupled with a precoding mode selection algorithm for each independently-decodable stream segment, which adjusts the downscaling factor according to scene characteristics, the utilized encoder, and the desired bitrate and encoding configuration. Our framework is compatible with all current and future codec and transport standards, as our deep precoding network structure is trained in conjunction with linear upscaling filters (e.g., the bilinear filter), which are supported by all web video players. Results with FHD and UHD content and widely-used AVC, HEVC and VP9 encoders show that coupling such standards with the proposed deep video precoding allows for 15% to 45% rate reduction under encoding configurations and bitrates suitable for video-on-demand adaptive streaming systems. The use of precoding can also lead to encoding complexity reduction, which is essential for cost-effective cloud deployment of complex encoders like H.265/HEVC and VP9.

Mohammad Jubran, Alhabib Abbas, Aaron Chadha et al.

Advanced video classification systems decode video frames to derive the necessary texture and motion representations for ingestion and analysis by spatio-temporal deep convolutional neural networks (CNNs). However, when considering visual Internet-of-Things applications, surveillance systems and semantic crawlers of large video repositories, the video capture and the CNN-based semantic analysis parts do not tend to be co-located. This necessitates the transport of compressed video over networks and incurs significant overhead in bandwidth and energy consumption, thereby significantly undermining the deployment potential of such systems. In this paper, we investigate the trade-off between the encoding bitrate and the achievable accuracy of CNN-based video classification models that directly ingest AVC/H.264 and HEVC encoded videos. Instead of retaining entire compressed video bitstreams and applying complex optical flow calculations prior to CNN processing, we only retain motion vector and select texture information at significantly-reduced bitrates and apply no additional processing prior to CNN ingestion. Based on three CNN architectures and two action recognition datasets, we achieve 11%-94% saving in bitrate with marginal effect on classification accuracy. A model-based selection between multiple CNNs increases these savings further, to the point where, if up to 7% loss of accuracy can be tolerated, video classification can take place with as little as 3 kbps for the transport of the required compressed video information to the system implementing the CNN models.

Aaron Chadha, Yiannis Andreopoulos

Adversarial discriminative domain adaptation (ADDA) is an efficient framework for unsupervised domain adaptation in image classification, where the source and target domains are assumed to have the same classes, but no labels are available for the target domain. We investigate whether we can improve performance of ADDA with a new framework and new loss formulations. Following the framework of semi-supervised GANs, we first extend the discriminator output over the source classes, in order to model the joint distribution over domain and task. We thus leverage on the distribution over the source encoder posteriors (which is fixed during adversarial training) and propose maximum mean discrepancy (MMD) and reconstruction-based loss functions for aligning the target encoder distribution to the source domain. We compare and provide a comprehensive analysis of how our framework and loss formulations extend over simple multi-class extensions of ADDA and other discriminative variants of semi-supervised GANs. In addition, we introduce various forms of regularization for stabilizing training, including treating the discriminator as a denoising autoencoder and regularizing the target encoder with source examples to reduce overfitting under a contraction mapping (i.e., when the target per-class distributions are contracting during alignment with the source). Finally, we validate our framework on standard domain adaptation datasets, such as SVHN and MNIST. We also examine how our framework benefits recognition problems based on modalities that lack training data, by introducing and evaluating on a neuromorphic vision sensing (NVS) sign language recognition dataset, where the source and target domains constitute emulated and real neuromorphic spike events respectively. Our results on all datasets show that our proposal competes or outperforms the state-of-the-art in unsupervised domain adaptation.

Aaron Chadha, Alhabib Abbas, Yiannis Andreopoulos

We investigate video classification via a two-stream convolutional neural network (CNN) design that directly ingests information extracted from compressed video bitstreams. Our approach begins with the observation that all modern video codecs divide the input frames into macroblocks (MBs). We demonstrate that selective access to MB motion vector (MV) information within compressed video bitstreams can also provide for selective, motion-adaptive, MB pixel decoding (a.k.a., MB texture decoding). This in turn allows for the derivation of spatio-temporal video activity regions at extremely high speed in comparison to conventional full-frame decoding followed by optical flow estimation. In order to evaluate the accuracy of a video classification framework based on such activity data, we independently train two CNN architectures on MB texture and MV correspondences and then fuse their scores to derive the final classification of each test video. Evaluation on two standard datasets shows that the proposed approach is competitive to the best two-stream video classification approaches found in the literature. At the same time: (i) a CPU-based realization of our MV extraction is over 977 times faster than GPU-based optical flow methods; (ii) selective decoding is up to 12 times faster than full-frame decoding; (iii) our proposed spatial and temporal CNNs perform inference at 5 to 49 times lower cloud computing cost than the fastest methods from the literature.

Aaron Chadha, Yiannis Andreopoulos

We investigate the problem of image retrieval based on visual queries when the latter comprise arbitrary regions-of-interest (ROI) rather than entire images. Our proposal is a compact image descriptor that combines the state-of-the-art in content-based descriptor extraction with a multi-level, Voronoi-based spatial partitioning of each dataset image. The proposed multi-level Voronoi-based encoding uses a spatial hierarchical K-means over interest-point locations, and computes a content-based descriptor over each cell. In order to reduce the matching complexity with minimal or no sacrifice in retrieval performance: (i) we utilize the tree structure of the spatial hierarchical K-means to perform a top-to-bottom pruning for local similarity maxima; (ii) we propose a new image similarity score that combines relevant information from all partition levels into a single measure for similarity; (iii) we combine our proposal with a novel and efficient approach for optimal bit allocation within quantized descriptor representations. By deriving both a Voronoi-based VLAD descriptor (termed as Fast-VVLAD) and a Voronoi-based deep convolutional neural network (CNN) descriptor (termed as Fast-VDCNN), we demonstrate that our Voronoi-based framework is agnostic to the descriptor basis, and can easily be slotted into existing frameworks. Via a range of ROI queries in two standard datasets, it is shown that the Voronoi-based descriptors achieve comparable or higher mean Average Precision against conventional grid-based spatial search, while offering more than two-fold reduction in complexity. Finally, beyond ROI queries, we show that Voronoi partitioning improves the geometric invariance of compact CNN descriptors, thereby resulting in competitive performance to the current state-of-the-art on whole image retrieval.

Francesco Renna, Joseph Doyle, Vasileios Giotsas et al.

Audio/visual recognition and retrieval applications have recently garnered significant attention within Internet-of-Things (IoT) oriented services, given that video cameras and audio processing chipsets are now ubiquitous even in low-end embedded systems. In the most typical scenario for such services, each device extracts audio/visual features and compacts them into feature descriptors, which comprise media queries. These queries are uploaded to a remote cloud computing service that performs content matching for classification or retrieval applications. Two of the most crucial aspects for such services are: (i) controlling the device energy consumption when using the service; (ii) reducing the billing cost incurred from the cloud infrastructure provider. In this paper we derive analytic conditions for the optimal coupling between the device energy consumption and the incurred cloud infrastructure billing. Our framework encapsulates: the energy consumption to produce and transmit audio/visual queries, the billing rates of the cloud infrastructure, the number of devices concurrently connected to the same cloud server, {the query volume constraint of each cluster of devices,} and the statistics of the query data production volume per device. Our analytic results are validated via a deployment with: (i) the device side comprising compact image descriptors (queries) computed on Beaglebone Linux embedded platforms and transmitted to Amazon Web Services (AWS) Simple Storage Service; (ii) the cloud side carrying out image similarity detection via AWS Elastic Compute Cloud (EC2) instances, with the AWS Auto Scaling being used to control the number of instances according to the demand.

Alhabib Abbas, Nikos Deligiannis, Yiannis Andreopoulos

We propose a novel vector aggregation technique for compact video representation, with application in accurate similarity detection within large video datasets. The current state-of-the-art in visual search is formed by the vector of locally aggregated descriptors (VLAD) of Jegou et. al. VLAD generates compact video representations based on scale-invariant feature transform (SIFT) vectors (extracted per frame) and local feature centers computed over a training set. With the aim to increase robustness to visual distortions, we propose a new approach that operates at a coarser level in the feature representation. We create vectors of locally aggregated centers (VLAC) by first clustering SIFT features to obtain local feature centers (LFCs) and then encoding the latter with respect to given centers of local feature centers (CLFCs), extracted from a training set. The sum-of-differences between the LFCs and the CLFCs are aggregated to generate an extremely-compact video description used for accurate video segment similarity detection. Experimentation using a video dataset, comprising more than 1000 minutes of content from the Open Video Project, shows that VLAC obtains substantial gains in terms of mean Average Precision (mAP) against VLAD and the hyper-pooling method of Douze et. al., under the same compaction factor and the same set of distortions.

Nikos Deligiannis, Joao F. C. Mota, George Smart et al.

Distributed desynchronization algorithms are key to wireless sensor networks as they allow for medium access control in a decentralized manner. In this paper, we view desynchronization primitives as iterative methods that solve optimization problems. In particular, by formalizing a well established desynchronization algorithm as a gradient descent method, we establish novel upper bounds on the number of iterations required to reach convergence. Moreover, by using Nesterov's accelerated gradient method, we propose a novel desynchronization primitive that provides for faster convergence to the steady state. Importantly, we propose a novel algorithm that leads to decentralized time-synchronous multichannel TDMA coordination by formulating this task as an optimization problem. Our simulations and experiments on a densely-connected IEEE 802.15.4-based wireless sensor network demonstrate that our scheme provides for faster convergence to the steady state, robustness to hidden nodes, higher network throughput and comparable power dissipation with respect to the recently standardized IEEE 802.15.4e-2012 time-synchronized channel hopping (TSCH) scheme.

Mohammad Ashraful Anam, Paul N. Whatmough, Yiannis Andreopoulos

Generic matrix multiplication (GEMM) and one-dimensional convolution/cross-correlation (CONV) kernels often constitute the bulk of the compute- and memory-intensive processing within image/audio recognition and matching systems. We propose a novel method to scale the energy and processing throughput of GEMM and CONV kernels for such error-tolerant multimedia applications by adjusting the precision of computation. Our technique employs linear projections to the input matrix or signal data during the top-level GEMM and CONV blocking and reordering. The GEMM and CONV kernel processing then uses the projected inputs and the results are accumulated to form the final outputs. Throughput and energy scaling takes place by changing the number of projections computed by each kernel, which in turn produces approximate results, i.e. changes the precision of the performed computation. Results derived from a voltage- and frequency-scaled ARM Cortex A15 processor running face recognition and music matching algorithms demonstrate that the proposed approach allows for 280%~440% increase of processing throughput and 75%~80% decrease of energy consumption against optimized GEMM and CONV kernels without any impact in the obtained recognition or matching accuracy. Even higher gains can be obtained if one is willing to tolerate some reduction in the accuracy of the recognition and matching applications.

Mohammad Ashraful Anam, Yiannis Andreopoulos

Convolution and cross-correlation are the basis of filtering and pattern or template matching in multimedia signal processing. We propose two throughput scaling options for any one-dimensional convolution kernel in programmable processors by adjusting the imprecision (distortion) of computation. Our approach is based on scalar quantization, followed by two forms of tight packing in floating-point (one of which is proposed in this paper) that allow for concurrent calculation of multiple results. We illustrate how our approach can operate as an optional pre- and post-processing layer for off-the-shelf optimized convolution routines. This is useful for multimedia applications that are tolerant to processing imprecision and for cases where the input signals are inherently noisy (error tolerant multimedia applications). Indicative experimental results with a digital music matching system and an MPEG-7 audio descriptor system demonstrate that the proposed approach offers up to 175% increase in processing throughput against optimized (full-precision) convolution with virtually no effect in the accuracy of the results. Based on marginal statistics of the input data, it is also shown how the throughput and distortion can be adjusted per input block of samples under constraints on the signal-to-noise ratio against the full-precision convolution.