Recent progress in face detection (including keypoint detection), and recognition is mainly being driven by (i) deeper convolutional neural network architectures, (ii) larger datasets. However, most of the large datasets are maintained private companies not publicly available. The academic computer vision community needs more varied to make further progress. In this paper, we introduce a new dataset, called UMDFaces, which has 367,888 annotated faces 8,277 subjects. We also evaluation...

Recent developments in deep convolutional neural networks (DCNNs) have shown impressive performance improvements on various object detection/recognition problems. This has been made possible due to the availability of large annotated data and a better understanding nonlinear mapping between images class labels, as well affordability powerful graphics processing units (GPUs). These learning also improved capabilities machines faces automatically executing tasks face detection, pose...

The availability of large annotated datasets and affordable computation power have led to impressive improvements in the performance convolutional neural networks (CNNs) on various face analysis tasks. In this paper, we describe a deep learning pipeline for unconstrained identification verification which achieves state-of-the-art several benchmark datasets. We provide design details modules involved automatic recognition: detection, landmark localization alignment,...

We present an approach for detecting human-object interactions (HOIs) in images, based on the idea that humans interact with functionally similar objects a manner. The proposed model is simple and efficiently uses data, visual features of human, relative spatial orientation human object, knowledge take part humans. provide extensive experimental validation our demonstrate state-of-the-art results HOI detection. On HICO-Det dataset method achieves gain over 2.5% absolute points mean average...

In this paper, targeted fooling of high performance image classifiers is achieved by developing two novel attack methods. The first method generates universal perturbations for target classes and the second specific perturbations. Extensive experiments are conducted on MNIST CIFAR10 datasets to provide insights about proposed algorithms show their effectiveness.

While the research community appears to have developed a consensus on methods of acquiring annotated data, design and training CNNs, many questions still remain be answered. In this paper, we explore following that are critical face recognition research: (i) Can train images expect systems work videos? (ii) Are deeper datasets better than wider datasets? (iii) Does adding label noise lead improvement in performance deep networks? (iv) Is alignment needed for recognition? We address these by...

Recent progress on action recognition has mainly focused RGB and optical flow features. In this paper, we approach the problem of joint-based recognition. Unlike other modalities, constellation joints their motion generate models with succinct human information for activity We present a new model recognition, which first extracts features from each joint separately through shared encoder before performing collective reasoning. Our selector module re-weights to select most discriminative...

Unconstrained face verification is a challenging problem owing to variations in pose, illumination, resolution of image, age, etc. This becomes even more complex when the subjects are actively trying deceive systems by wearing disguise. The under consideration here identify subject disguises and reject impostors look like interest. In this paper we present DCNN-based approach for recognizing people picking out impostors. We train two different networks on large dataset comprising still...

In recent years, the performance of face verification and recognition systems based on deep convolutional neural networks (DCNNs) has significantly improved. A typical pipeline for includes training a network subject classification with softmax loss, using penultimate layer output as feature descriptor, generating cosine similarity score given pair images or videos. The loss function does not optimize features to have higher positive pairs lower negative pairs, which leads gap. this paper,...

We present a new formulation for structured information extraction (SIE) from visually rich documents. address the limitations of existing IOB tagging and graph-based formulations, which are either overly reliant on correct ordering input text or struggle with decoding complex graph. Instead, motivated by anchor-based object detectors in computer vision, we represent an entity as anchor word bounding box, linking association between words. This is more robust to ordering, maintains compact...

As deep networks become increasingly accurate at recognizing faces, it is vital to understand how these process faces. While are solely trained recognize identities, they also contain face related information such as sex, age, and pose of the even when not learn attributes. We introduce expressivity a measure much feature vector informs us about an attribute, where can be from internal or final layers network. Expressivity computed by second neural network whose inputs features The output...

We present a method of estimating the number people in high density crowds from still images. The estimates counts by fusing information multiple sources. Most existing work on crowd counting deals with very small (tens individuals) and use temporal videos. Our uses only images to estimate (hundreds thousands individuals). At this scale, we cannot rely one set features for count estimation. We, therefore, sources, viz. interest points (SIFT), Fourier analysis, wavelet decomposition, GLCM low...

A core component of the recent success self-supervised learning is cropping data augmentation, which selects sub-regions an image to be used as positive views in loss. The underlying assumption that randomly cropped and resized regions a given share information about objects interest, learned representation will capture. This mostly satisfied datasets such ImageNet where there large, centered object, highly likely present random crops full image. However, other OpenImages or COCO, are more...

We present a new formulation for structured information extraction (SIE) from visually rich documents. It aims to address the limitations of existing IOB tagging or graph-based formulations, which are either overly reliant on correct ordering input text struggle with decoding complex graph. Instead, motivated by anchor-based object detectors in vision, we represent an entity as anchor word and bounding box, linking association between words. This is more robust ordering, maintains compact...

A human pose often conveys not only the configuration of body parts, but also implicit predictive information about ensuing motion. This dynamic can benefit vision applications which lack explicit motion cues. The visual system easily perceive in still images. However, computational algorithms to infer and utilize it computer are limited. In this paper, we propose a probabilistic framework associated with pose. inference problem is posed as nonparametric density estimation on non-Euclidean...