Standard video codecs are rate-distortion optimization machines, where distortion is typically quantified using PSNR versus the source. However, it is now widely accepted that increasing PSNR does not necessarily translate to better visual quality. In this paper, a better balance between perception and fidelity is targeted, in order to provide for significant rate savings over state-of-the-art standards-based video codecs. Specifically, pre- and post-processing neural networks are proposed that enhance the coding efficiency of standard video codecs when bench-marked with an array of well-established perceptual quality scores. These "neural wrapper" elements are end-to-end trained with a neural codec module serving as a differentiable proxy for standard video codecs. The codec proxy is jointly optimized with the pre- and post components via a novel two-phase pretraining strategy and end-to-end iterative refinement with stop-gradient. This allows the neural pre- and postprocessor to learn to embed, remove and recover information in a codec-aware manner, thus improving its rate-quality performance. A single neural-wrapper model is thereby established and used for the entire rate-quality curve without needing any downscaling or upscaling. The trained model is tested with the AV1 and VVC standard codecs via an array of well-established objective quality scores (SSIM, MS-SSIM, VMAF, AVQT), as well as mean opinion scores (MOS) derived from ITU-T P.910 subjective testing. Experimental results show that the proposed approach improves all quality scores, with -18.5% average Bjontegaard Delta-rate (BD-rate) saving over all objective scores and MOS improvement over both standard codecs. This illustrates the significant potential of neural wrapper components over standards-based video coding.

Video processing algorithms like video upscaling, denoising, and compression are now increasingly optimized for perceptual quality metrics instead of signal distortion. This means that they may score well for metrics like video multi-method assessment fusion (VMAF), but this may be because of metric overfitting. This imposes the need for costly subjective quality assessments that cannot scale to large datasets and large parameter explorations. We propose a methodology that fuses multiple quality metrics based on small scale subjective testing in order to unlock their use at scale for specific application domains of interest. This is achieved by employing pseudo-random sampling of the resolution, quality range and test video content available, which is initially guided by quality metrics in order to cover the quality range useful to each application. The selected samples then undergo a subjective test, such as ITU-T P.910 absolute categorical rating, with the results of the test postprocessed and used as the means to derive the best combination of multiple objective metrics using support vector regression. We showcase the benefits of this approach in two applications: video encoding with and without perceptual preprocessing, and deep video denoising & upscaling of compressed content. For both applications, the derived fusion of metrics allows for a more robust alignment to mean opinion scores than a perceptually-uninformed combination of the original metrics themselves. The dataset and code is available at https://github.com/isize-tech/VideoQualityFusion.

We introduce the concept of rate-aware deep perceptual preprocessing (DPP) for video encoding. DPP makes a single pass over each input frame in order to enhance its visual quality when the video is to be compressed with any codec at any bitrate. The resulting bitstreams can be decoded and displayed at the client side without any post-processing component. DPP comprises a convolutional neural network that is trained via a composite set of loss functions that incorporates: (i) a perceptual loss based on a trained no-reference image quality assessment model, (ii) a reference-based fidelity loss expressing L1 and structural similarity aspects, (iii) a motion-based rate loss via block-based transform, quantization and entropy estimates that converts the essential components of standard hybrid video encoder designs into a trainable framework. Extensive testing using multiple quality metrics and AVC, AV1 and VVC encoders shows that DPP+encoder reduces, on average, the bitrate of the corresponding encoder by 11%. This marks the first time a server-side neural processing component achieves such savings over the state-of-the-art in video coding.

Several groups worldwide 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 H.264/AVC, H.265/HEVC, VVC, Google VP9 and AOMedia AV1, AV2, 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 when considering the fact that the video content industry and hardware manufacturers are expected to remain committed to supporting 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. Extensive evaluation on FHD (1080p) and UHD (2160p) content and with widely-used H.264/AVC, H.265/HEVC and VP9 encoders, as well as a preliminary evaluation with the current test model of VVC (v.6.2rc1), shows that coupling such standards with the proposed deep video precoding allows for 8% to 52% 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, VP9 and VVC, especially when considering the prominence of high-resolution adaptive video streaming.

Advanced video classification systems decode video frames to derive 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% savings 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 kb/s for the transport of the required compressed video information to the system implementing the CNN models.

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 appearancebased 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).

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. While ADDA has already achieved better training efficiency and competitive accuracy on image classification in comparison to other adversarial based methods, we investigate whether we can improve its performance 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 datasets like MNIST, USPS, SVHN, MNIST-M and Office-31. We additionally examine how the proposed framework benefits recognition problems based on sensing modalities that lack training data. This is realized by introducing and evaluating on a neuromorphic vision sensing (NVS) sign language recognition dataset, where the source domain constitutes emulated neuromorphic spike events converted from conventional pixel-based video and the target domain is experimental (real) spike events from an NVS camera. Our results on all datasets show that our proposal is both simple and efficient, as it competes or outperforms the state-of-the-art in unsupervised domain adaptation, such as DIFA and MCDDA, whilst offering lower complexity than other recent adversarial methods.

Neuromorphic vision sensing (NVS) hardware is now gaining traction as a low-power/high-speed visual sensing technology that circumvents the limitations of conventional active pixel sensing (APS) cameras. While object detection and tracking models have been investigated in conjunction with NVS, there is currently little work on NVS for higher-level semantic tasks, such as action recognition. Contrary to recent work that considers homogeneous transfer between flow domains (optical flow to motion vectors), we propose to embed an NVS emulator into a multi-modal transfer learning framework that carries out heterogeneous transfer from optical flow to NVS. The potential of our framework is showcased by the fact that, for the first time, our NVS-based results achieve comparable action recognition performance to motion-vector or optical-flow based methods (i.e., accuracy on UCF-101 within 8.8% of I3D with optical flow), with the NVS emulator and NVS camera hardware offering 3 to 6 orders of magnitude faster frame generation (respectively) compared to standard Brox optical flow. Beyond this significant advantage, our CNN processing is found to have the lowest total GFLOP count against all competing methods (up to 7.7 times complexity saving compared to I3D with optical flow).

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.

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 data sets shows that the proposed approach is competitive with the best two-stream video classification approaches found in the literature. At the same time: 1) a CPU-based realization of our MV extraction is over 977 times faster than GPU-based optical flow methods; 2) selective decoding is up to 12 times faster than full-frame decoding; and 3) our proposed spatial and temporal CNNs perform inference at 5 to 49 times lower cloud computing cost than the fastest methods from the literature.

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 compressed video content 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 that ingests AVC/H.264 encoded videos. Instead of entire compressed video bitstreams, we only retain motion vector and selected texture information at significantly reduced bitrates. Based on two CNN architectures and two action recognition datasets, we achieve 38%-59% saving in bitrate with marginal impact in classification accuracy. A simple rate-based selection between the two CNNs shows that even further bitrate savings are possible with graceful degradation in accuracy. This may allow for rate/accuracy-optimized CNN-based video classification over networks.

We investigate video classification via a 3D deep convolutional neural network (CNN) that directly ingests compressed bitstream information. This idea is based on the observation that video macroblock (MB) motion vectors (that are very compact and directly available from the compressed bitstream) are inherently capturing local spatio-temporal changes in each video scene. Our results on two standard video datasets show that our approach outperforms pixel-based approaches and remains within 7 percentile points from the best classification results reported by highly-complex optical-flow & deep-CNN methods. At the same time, a CPU-based realization of our approach is found to be more than 2500 times faster in the motion extraction in comparison to GPU-based optical flow methods and also offers 2 to 3.4-fold reduction in the utilized deep CNN weights compared to recent architectures. This indicates that deep learning based on compressed video bitstream information may allow for advanced video classification to be deployed in very large datasets using commodity CPU hardware. Source code is available at http://www.github.com/mvcnn.

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 multilevel, Voronoi-based spatial partitioning of each dataset image. The proposed multilevel 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: 1) we utilize the tree structure of the spatial hierarchical K-means to perform top-to-bottom pruning for local similarity maxima; 2) we propose a new image similarity score that combines relevant information from all partition levels into a single measure for similarity; 3) 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 (called Fast-VVLAD) and a Voronoi-based deep convolutional neural network (CNN) descriptor (called 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 twofold 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.