Pose-tracking is an important problem that requires identifying unique human pose-instances and matching them temporally across different frames in a video. However, existing pose-tracking methods are unable to accurately model temporal relationships and require significant computation, often computing the tracks offline. We present an efficient multi-person pose-tracking method, KeyTrack that only relies on keypoint information without using any RGB or optical flow to locate and track human keypoints in real-time. KeyTrack is a top-down approach that learns spatio-temporal pose relationships by modeling the multi-person pose-tracking problem as a novel Pose Entailment task using a Transformer based architecture. Furthermore, KeyTrack uses a novel, parameter-free, keypoint refinement technique that improves the keypoint estimates used by the Transformers. We achieve state-of-the-art results on PoseTrack'17 and PoseTrack'18 benchmarks while using only a fraction of the computation used by most other methods for computing the tracking information.

We propose a sequential variational autoencoder to learn disentangled representations of sequential data (e.g., videos and audios) under self-supervision. Specifically, we exploit the benefits of some readily accessible supervision signals from input data itself or some off-the-shelf functional models and accordingly design auxiliary tasks for our model to utilize these signals. With the supervision of the signals, our model can easily disentangle the representation of an input sequence into static factors and dynamic factors (i.e., time-invariant and time-varying parts). Comprehensive experiments across videos and audios verify the effectiveness of our model on representation disentanglement and generation of sequential data, and demonstrate that, our model with self-supervision performs comparable to, if not better than, the fully-supervised model with ground truth labels, and outperforms state-of-the-art unsupervised models by a large margin.

The rich and accessible labeled data fueled the revolutionary successes of deep learning in object recognition. However, recognizing objects of novel classes with limited supervision information provided, i.e., Novel Object Recognition (NOR), remains a challenging task. We identify in this paper two key factors for the success of NOR that previous approaches fail to simultaneously guarantee. The first is producing discriminative feature representations for images of novel classes, and the second is generating a flexible classifier readily adapted to novel classes provided with limited supervision signals. To secure both key factors, we propose a framework which decouples a deep classification model into a feature extraction module and a classification module. We learn the former to ensure feature discriminability with a standard multi-class classification task by fully utilizing the competing information among all classes within a training set, and learn the latter to secure adaptability by training a meta-learner network which generates classifier weights whenever provided with minimal supervision information of target classes. Extensive experiments on common benchmark datasets in the settings of both zero-shot and few-shot learning demonstrate our method achieves state-of-the-art performance.

Developing conditional generative models for text-to-video synthesis is an extremely challenging yet an important topic of research in machine learning. In this work, we address this problem by introducing Text-Filter conditioning Generative Adversarial Network (TFGAN), a conditional GAN model with a novel multi-scale text-conditioning scheme that improves text-video associations. By combining the proposed conditioning scheme with a deep GAN architecture, TFGAN generates high quality videos from text on challenging real-world video datasets. In addition, we construct a synthetic dataset of text-conditioned moving shapes to systematically evaluate our conditioning scheme. Extensive experiments demonstrate that TFGAN significantly outperforms existing approaches, and can also generate videos of novel categories not seen during training.

Human actions often involve complex interactions across several inter-related objects in the scene. However, existing approaches to fine-grained video understanding or visual relationship detection often rely on single object representation or pairwise object relationships. Furthermore, learning interactions across multiple objects in hundreds of frames for video is computationally infeasible and performance may suffer since a large combinatorial space has to be modeled. In this paper, we propose to efficiently learn higher-order interactions between arbitrary subgroups of objects for fine-grained video understanding. We demonstrate that modeling object interactions significantly improves accuracy for both action recognition and video captioning, while saving more than 3-times the computation over traditional pairwise relationships. The proposed method is validated on two large-scale datasets: Kinetics and ActivityNet Captions. Our SINet and SINet-Caption achieve state-of-the-art performances on both datasets even though the videos are sampled at a maximum of 1 FPS. To the best of our knowledge, this is the first work modeling object interactions on open domain large-scale video datasets, and we additionally model higher-order object interactions which improves the performance with low computational costs.

Applications that use learning and classification algorithms operate on large amounts of unstructured data, and have stringent performance constraints. For such applications, the performance of general purpose processors scales poorly with data size because of their limited support for fine-grained parallelism and absence of software-managed caches. The large intermediate data in these applications also limits achievable performance on many-core processors such as GPUs. To accelerate such learning applications, we present a programmable accelerator that can execute multiple learning and classification algorithms. To architect such an accelerator, we profile five representative workloads, and find that their computationally intensive portions can be formulated as matrix or vector operations generating large amounts of intermediate data, which are then reduced by a secondary operation such as array ranking, finding max/min and aggregation. Our proposed accelerator, called MAPLE, has hundreds of simple processing elements (PEs) laid out in a two-dimensional grid, with two key features. First, it uses dynamic in-memory processing where on-chip memory blocks perform the secondary reduction operations. Second, MAPLE uses banked off-chip memory, and organizes its PEs into independent groups each with its own off-chip memory bank. These two features allow MAPLE to scale its performance with data size. We also present an Atom based energy-efficient heterogeneous system with MAPLE as the accelerator that satisfies the application’s performance requirements at a lower system power. This article describes the MAPLE architecture, explores its design space with a simulator, illustrates how to automatically map application kernels to the hardware, and presents its performance improvement and energy benefits over classic server-based implementations. We implement a 512-PE FPGA prototype of MAPLE and find that it is 1.5-10x faster than a 2.5 GHz quad-core Xeon processor despite running at a modest 125 MHz clock rate. With MAPLE connected to a 1.6GHz dual-core Atom, we show an energy improvement of 38-84% over the Xeon server coupled to a 1.3 GHz 240 core Tesla GPU.

For learning and classification workloads that operate on large amounts of unstructured data with stringent performance constraints, general purpose processor performance scales poorly with data size. In this paper, we present a programmable accelerator for this workload domain. To architect the accelerator, we profile five representative workloads, and find that their computationally intensive portions can be formulated as matrix or vector operations generating large amounts of intermediate data, which are then reduced by a secondary operation such as array ranking, finding max/min and aggregation. The proposed accelerator, called MAPLE, has hundreds of simple processing elements (PEs) laid out in a two-dimensional grid, with two key features. First, it uses in-memory processing where on-chip memory blocks perform the secondary reduction operations. By doing so, the intermediate data are dynamically processed and never stored or sent off-chip. Second, MAPLE uses banked off-chip memory, and organizes its PEs into independent groups each with its own off-chip memory bank. These two features together allow MAPLE to scale its performance with data size. This paper describes the MAPLE architecture, explores its design space with a simulator, and illustrates how to automatically map application kernels to the hardware. We also implement a 512-PE FPGA prototype of MAPLE and find that it is 1.5-10x faster than a 2.5 GHz quad-core Xeon processor despite running at a modest 125 MHz.

We present a massively parallel coprocessor for accelerating Convolutional Neural Networks (CNNs), a class of important machine learning algorithms. The coprocessor functional units, consisting of parallel 2D convolution primitives and programmable units performing sub-sampling and non-linear functions specific to CNNs, implement a ldquometa-operatorrdquo to which a CNN may be compiled to. The coprocessor is serviced by distributed off-chip memory banks with large data bandwidth. As a key feature, we use low precision data and further increase the effective memory bandwidth by packing multiple words in every memory operation, and leverage the algorithmpsilas simple data access patterns to use off-chip memory as a scratchpad for intermediate data, critical for CNNs. A CNN is mapped to the coprocessor hardware primitives with instructions to transfer data between the memory and coprocessor. We have implemented a prototype of the CNN coprocessor on an off-the-shelf PCI FPGA card with a single Xilinx Virtex5 LX330T FPGA and 4 DDR2 memory banks totaling 1 GB. The coprocessor prototype can process at the rate of 3.4 billion multiply accumulates per second (GMACs) for CNN forward propagation, a speed that is 31x faster than a software implementation on a 2.2 GHz AMD Opteron processor. For a complete face recognition application with the CNN on the coprocessor and the rest of the image processing tasks on the host, the prototype is 6-10times faster, depending on the host-coprocessor bandwidth.

We present a massively parallel FPGA-based coprocessor for Support Vector Machines (SVMs), a machine learning algorithm whose applications include recognition tasks such as learning scenes, situations and concepts, and reasoning tasks such as analyzing the recognized scenes and semantics. The coprocessor architecture, targeted at both SVM training and classification, is based on clusters of vector processing elements (VPEs) operating in single-instruction multiple data (SIMD) mode to take advantage of large amounts of data parallelism in the application. We use the FPGA's DSP elements as parallel multiply-accumulators (MACs), a core computation in SVMs. A key feature of the architecture is that it is customized to low precision arithmetic which permits one DSP unit to perform two or more MACs in parallel. Low precision also reduces the required number of parallel off-chip memory accesses by packing multiple data words on the FPGA-memory bus. We have built a prototype using an off-the-shelf PCI-based FPGA card with a Xilinx Virtex 5 FPGA and 1 GB DDR2 memory. For SVM training, we observe application-level end-to-end computation speeds of over 9 billion multiply-accumulates per second (GMACs). For SVM classification, using data packing, the application speed increases to 14 GMACs. The FPGA-based system is about 20times faster than a dual Opteron 2.2 GHz processor CPU, and dissipates around 10 W of power.

A mainstay in cancer diagnostics is the classification or grading of cell nuclei based on their appearance. While the analysis of cytological samples has been automated successfully for a long time, the complexity of histological tissue samples has prevented a reliable classification with machine vision techniques. We approach this complex problem in multiple stages, analyzing first image quality, staining quality, and tissue appearance, before segmenting nuclei and finally classifying or grading areas of tissue. The key step is the training of a classifier to judge the nuclei segmentation quality. Using active learning techniques, we train this classifier to identify problems in the image as well as weaknesses of the image analysis tools. This way we obtain robust nuclear segmentation allowing precise measurements of features that can be used safely for classification. We validate our findings on several hundred cases of breast cancer, demonstrating that automatic pleomorphism grading is possible with high accuracy. This technique can provide a stable and objective basis for what has been a subjective process that suffers from low reproducibility.

We present a new, massively parallel architecture for accelerating machine learning algorithms, based on arrays of variable-resolution arithmetic vector processing elements (VPE). Groups of VPEs operate in SIMD (single instruction multiple data) mode, and each group is connected to an independent memory bank. In this way memory bandwidth scales with the number of VPE, and the main data flows are local, keeping power dissipation low. With 256 VPEs, implemented on two FPGA (field programmable gate array) chips, we obtain a sustained speed of 19 GMACS (billion multiply-accumulate per sec.) for SVM training, and 86 GMACS for SVM classification. This performance is more than an order of magnitude higher than that of any FPGA implementation reported so far. The speed on one FPGA is similar to the fastest speeds published on a Graphics Processor for the MNIST problem, despite a clock rate of the FPGA that is six times lower. High performance at low clock rates makes this massively parallel architecture particularly attractive for embedded applications, where low power dissipation is critical. Tests with Convolutional Neural Networks and other learning algorithms are under way now.

We describe an algorithm for support vector machines (SVM) that can be parallelized efficiently and scales to very large problems with hundreds of thousands of training vectors. Instead of analyzing the whole training set in one optimization step, the data are split into subsets and optimized separately with multiple SVMs. The partial results are combined and filtered again in a ‘Cascade’ of SVMs, until the global optimum is reached. The Cascade SVM can be spread over multiple processors with minimal communication overhead and requires far less memory, since the kernel matrices are much smaller than for a regular SVM. Convergence to the global optimum is guaranteed with multiple passes through the Cascade, but already a single pass provides good generalization. A single pass is 5x – 10x faster than a regular SVM for problems of 100,000 vectors when implemented on a single processor. Parallel implementations on a cluster of 16 processors were tested with over 1 million vectors (2-class problems), converging in a day or two, while a regular SVM never converged in over a week.

Lifelike talking faces for interactive services are an exciting new modality for man-machine interactions. Recent developments in speech synthesis and computer animation enable the real-time synthesis of faces that look and behave like real people, opening opportunities to make interactions with computers more like face-to-face conversations. This paper focuses on the technologies for creating lifelike talking heads, illustrating the two main approaches: model-based animations and sample-based animations. The traditional model-based approach uses three-dimensional wire-frame models, which can be animated from high-level parameters such as muscle actions, lip postures, and facial expressions. The sample-based approach, on the other hand, concatenates segments of recorded videos, instead of trying to model the dynamics of the animations in detail. Recent advances in image analysis enable the creation of large databases of mouth and eye images, suited for sample-based animations. The sample-based approach tends to generate more naturally looking animations at the expense of a larger size and less flexibility than the model-based animations. Beside lip articulation, a talking head must show appropriate head movements, in order to appear natural. We illustrate how such "visual prosody" is analyzed and added to the animations. Finally, we present four applications where the use of face animation in interactive services results in engaging user interfaces and an increased level of trust between user and machine. Using an RTP-based protocol, face animation can be driven with only 800 bits/s in addition to the rate for transmitting audio.

Concatenative visual speech synthesis selects frames from a large recorded video database of mouth shapes to generate photo-realistic talking head sequences. The synthesized sequence must exhibit precise lip-sound synchronization and smooth articulation. The selection process for finding the best lip shapes has been computationally expensive [1], limiting the speed of the synthesis to far less than real time. In this paper, we propose a rapid unit selection approach based on triphone units. Experiments show that this algorithm can make the synthesis, excluding the rendering, 50 times faster than real-time on a standard desktop PC. We also developed a metric to test the quality of the synthesis objectively, and show that this measurement is consistent with subjective measurement results.

In this paper we discuss a simple method to improve the lighting conditions of a scene or video sequence. We focus on modifying the intensities of existing light sources and on inserting additional virtual lights into a scene. We target video conferencing applications where typical lighting conditions (i e the average office or home environment) are poor. Our model is much simpler and is a departure from approaches such as re-lighting, developed in computer-augmented reality (CAR) where the scene is first reconstructed geometrically by means of computer vision techniques We rely on a well-known illumination equation that takes into account light attenuation, plus lambertian and specular reflection Our first scenario models a 3D space in which the image or video frame lies on a 2D plane. Virtual lights are placed in the 3D space and illuminate the 2D image We then propose a refined model for talking head video sequences where the talking head is detected, tracked and its surface is modeled with a 3D ellipsoid. Results show that convincing lighting improvements can be achieved in both scenarios.

This paper describes techniques for extracting bitmaps of facial parts from videos of a talking person. The goal is to synthesize photo-realistic talking heads of high quality that show picture-perfect appearance and realistic head movements with good lip-sound synchronization. For the synthesis of a talking head, bitmaps of facial parts are combined to form whole heads and then sequences of such images are integrated with audio from a text-to-speech synthesizer. For a seamless integration of facial parts into an animation, their shape and visual appearance must be known with high accuracy. The recognition system has to find not only the locations of facial features, but must also be able to determine the head's orientation and recognize the facial expressions. Our face recognition proceeds in multiple steps, each with an increased precision. Using motion, color and shape information, the head's position and the location of the main facial features are determined first. Then smaller areas are searched with matched filters, in order to identify specific facial features with high precision. From this information a head's 3D orientation is calculated. Facial parts are cut from the image and, using the head's orientation, are warped into bitmaps with 'normalized' orientation and scale.

This paper describes a system for creating a photo-realistic model of the human head that can be animated and lip-synched from phonetic transcripts of text. Combined with a state-of-the-art text-to-speech synthesizer (TTS), it generates video animations of talking heads that closely resemble real people. To obtain a naturally looking head, we choose a "data-driven" approach. We record a talking person and apply image recognition to extract automatically bitmaps of facial parts. These bitmaps are normalized and parameterized before being entered into a database. For synthesis, the TTS provides the audio track, as well as the phonetic transcript from which trajectories in the space of parameterized bitmaps are computed for all facial parts. Sampling these trajectories and retrieving the corresponding bitmaps from the database produces animated facial parts. These facial parts are then projected and blended onto an image of the whole head using its pose information. This talking head model can produce new never recorded speech of the person who was originally recorded. Talking-head animations of this type are useful as a front-end for agents and avatars in multimedia applications such as virtual operators, virtual announcers, help desks, educational, and expert systems.

The paper describes a system that generates photo-realistic video animations of talking heads. First the system derives head models from existing video footage using image recognition techniques. It locates, extracts and labels facial parts such as mouth, eyes, and eyebrows into a compact library. Then, using these face models and a text-to-speech synthesizer, it synthesizes new video sequences of the head where the lips are in synchrony with the accompanying soundtrack. Emotional cues and conversational signals are produced by combining head movements, raising eyebrows, wide open eyes, etc. with the mouth animation. For these animations to be believable, care has to be taken aligning the facial parts so that they blend smoothly into each other and produce seamless animations. Our system uses precise multi channel facial recognition techniques to track facial parts, and it derives the exact 3D position of the head, enabling the automatic extraction of normalized face parts. Such talking head animations are useful because they generally increase intelligibility of the human machine interface in applications where content needs to be narrated to the user, such as educative software.

We propose the use of discriminative training by means of the generalized probabilistic descent (GPB) algorithm to estimate hidden Markov model (HMM) stream exponents for audio-visual speech recognition. Synchronized audio and visual features are used to respectively train audio-only and visual-only single-stream HMMs of identical topology by maximum likelihood. A two-stream HMM is then obtained by combining the two single-stream HMMs and introducing exponents that weigh the log-likelihood of each stream. We present the GPD algorithm for stream exponent estimation, consider a possible initialization, and apply it to the single speaker connected letters task of the AT&T bimodal database. We demonstrate the superior performance of the resulting multi-stream HMM to the audio-only, visual-only, and audio-visual single-stream HMMs.

This paper investigates the use of Fisher-Rao (1965) linear discriminant analysis (LDA) as a means of visual feature extraction for hidden Markov model based automatic speechreading. For every video frame, a three-dimensional region of interest containing the speaker's mouth over a sequence of adjacent frames is lexicographically arranged into a data vector. Such vectors are then projected onto the space of the most discriminant "eigensequences", estimated by means of LDA on a training set of image sequence vectors, labeled from a set of a-priori chosen classes. The resulting projections, as well as their first and second derivatives over time, are used as features for automatic speechreading. The proposed method is applied to single-speaker, multi-speaker, and speaker-independent visual-only recognition tasks, consistently outperforming principal component analysis and discrete wavelet transform based visual features. Specific issues relevant to LDA are also discussed, namely, class selection, automatic data class labelling, and dimensionality reduction prior to LDA.

We designed a modular system using a combination of shape analysis, color segmentation and motion information for locating reliably heads and faces of different sizes and orientations in complex images. The first of the system's three channels does a shape analysis on gray-level images to determine the location of individual facial features as well as the outlines of heads. In the second channel the color space is analyzed with a clustering algorithm to find areas of skin colors. The color space is first calibrated, using the results from the other channels. In the third channel motion information is extracted from frame differences. Head outlines are determined by analyzing the shapes of areas with large motion vectors. All three channels produce lists of shapes, each marking an area of the image where a facial feature or a part of the outline of a head may be present. Combinations of such shapes are evaluated with n-gram searches to produce a list of likely head positions and the locations of facial features. We tested the system for tracking faces of people sitting in front of terminals and video phones and used it to track people entering through a doorway.

The robust acquisition of facial features needed for visual speech processing is fraught with difficulties which greatly increase the complexity of the machine vision system. This system must extract the inner lip contour from facial images with variations in pose, lighting, and facial hair. This paper describes a face feature acquisition system with robust performance in the presence of extreme lighting variations and moderate variations in pose. Furthermore, system performance is not degraded by facial hair or glasses. To find the position of a face reliably we search the whole image for facial features. These features are then combined and tests are applied, to determine whether any such combination actually belongs to a face. In order to find where the lips are, other features of the face, such as the eyes, must be located as well. Without this information it is difficult to reliably find the mouth in a complex image. Just the mouth by itself is easily missed or other elements in the image can be mistaken for a mouth. If camera position can be constrained to allow the nostrils to be viewed, then nostril tracking is used to both reduce computation and provide additional robustness. Once the nostrils are tracked from frame to frame using a tracking window the mouth area can be isolated and normalized for scale and rotation. A mouth detail analysis procedure is then used to estimate the inner lip contour and teeth and tongue regions. The inner lip contour and head movements are then mapped to synthetic face parameters to generate a graphical talking head synchronized with the original human voice. This information can also be used as the basis for visual speech features in an automatic speechreading system. Similar features were used in our previous automatic speechreading systems.

Two ANNA neural-network chips are integrated on a 6U VME board, to serve as a high-speed platform for a wide variety of algorithms used in neural-network applications as well as in image analysis. The system can implement neural networks of variable sizes and architectures, but can also be used for filtering and feature extraction tasks that are based on convolutions. The board contains a controller implemented with field programmable gate arrays (FPGA's), memory, and bus interfaces, all designed to support the high compute power of the ANNA chips. This new system is designed for maximum speed and is roughly 10 times faster than a previous board. The system has been tested for such tasks as text location, character recognition, and noise removal as well as for emulating cellular neural networks (CNN's). A sustained speed of up to two billion connections per second (GC/s) and a recognition speed of 1000 characters per second has been measured.

We utilize speech information to improve the quality of audio-visual communications such as video telephony and videoconferencing. We show that the marriage of speech analysis and image processing can solve problems related to lip synchronization. We present a technique called speech-assisted frame-rate conversion, and apply it to coding of talking head video. Demonstration sequences are presented. Extensions and other applications are outlined.

A video codec typically-skips frames to satisfy the bit rate constraint. This results in a jerky motion and a loss of lip synchronization in sequences of talking persons. We propose a frame interpolation technique to solve these problems. The main components of this technique include: foreground/background segmentation, mesh-based frame interpolation, and image analysis/synthesis for mouth movements. This technique generates smooth head motion and renders lip motion that is synchronized with the voice of the person.

We describe an image analysis system for handling complex and noisy images of forms and bank documents, such as business checks, personal checks, or bank deposits. Some of these document types have no standardized layout, requiring a careful analysis of the whole image, to find out where the relevant information, for example the courtesy amount, is located. Each element in the image is first classified as being part of machine printed text, handwritten text, or as being a graphical element, such as a line. To obtain a reliable identification of these different elements under noisy conditions, a set of templates is scanned over the image, extracting such elements as strokes, line end stops and corners. From this representation a quick and robust analysis of the image's content is possible to identify the different parts. Once a set of candidate subimages has been found, they are sent to a field recognition system. We describe an example of one such system, which locates and reads courtesy amounts on US checks.

We utilize speech information to improve the quality of audio-visual communications such as videotelephony and videoconferencing. In particular, the marriage of speech analysis and image processing can solve problems related to lip synchronization. We present a technique called speech-assisted frame-rate conversion. Demonstration sequences are presented. Other applications, including speech-assisted video coding, are outlined.>

The authors applied an analog neural net chip to several machine vision tasks, among them: locate the address blocks on mail pieces, find handwritten text on checks, and discriminate between handwritten and machine printed characters. The chip, operating as a co-processor of a workstation, provides a speed-up of about a factor of 1000, compared with the workstation. The computation speed achieved lies between one and ten billion multiply-accumulates per second. The neural net chip is based on building blocks, 'neurons', that can be arranged in various network architectures. The data flow is optimized for implementing large, structured neural nets, and is also suited for any task where signals are to be convolved with many kernels. Some of the networks are trained on the neural net chip with a weight-perturbation learning algorithm that was adapted to work with the coarse quantization of the weights and the states in the chip.>

Hardware architectures for character recognition are discussed, and choices for possible circuits are outlined. An advanced (and working) reconfigurable neural-net chip that mixes analog and digital processing is described. It is found that different approaches to image recognition often lead to neural-net architectures that have limited connectivity and repeated use of the same set of weights. This architecture is ideal for time-multiplexing (a combined parallel-series processing) on hardware systems that would be too small to evaluate the entire network in parallel. To make this process efficient, a chip needs to have shift registers to format the input data and additional registers to store intermediate results. Within this framework, it is possible to design chips that have broad utility, large connection capacity, and high speed. This was demonstrated by a new chip with 32000 reconfigurable connections