Training a neural network is a monolithic endeavor, akin to carving knowledge into stone: once the process is completed, editing the knowledge in a network is hard, since all information is distributed across the network's weights. We here explore a simple, compelling alternative by marrying the representational power of deep neural networks with the flexibility of a database. Decomposing the task of image classification into image similarity (from a pre-trained embedding) and search (via fast nearest neighbor retrieval from a knowledge database), we build on well-established components to construct a simple and flexible visual memory that has the following key capabilities: (1.) The ability to flexibly add data across scales: from individual samples all the way to entire classes and billion-scale data; (2.) The ability to remove data through unlearning and memory pruning; (3.) An interpretable decision-mechanism on which we can intervene to control its behavior. Taken together, these capabilities comprehensively demonstrate the benefits of an explicit visual memory. We hope that it might contribute to a conversation on how knowledge should be represented in deep vision models---beyond carving it in "stone" weights.

By optimizing the rate-distortion-realism trade-off, generative compression approaches produce detailed, realistic images, even at low bit rates, instead of the blurry reconstructions produced by rate-distortion optimized models. However, previous methods do not explicitly control how much detail is synthesized, which results in a common criticism of these methods: users might be worried that a misleading reconstruction far from the input image is generated. In this work, we alleviate these concerns by training a decoder that can bridge the two regimes and navigate the distortion-realism tradeoff. From a single compressed representation, the receiver can decide to either reconstruct a low mean squared error reconstruction that is close to the input, a realistic reconstruction with high perceptual quality, or anything in between. With our method, we set a new state-of-the-art in distortion-realism, pushing the frontier of achievable distortion-realism pairs, i.e., our method achieves better distortions at high realism and better realism at low distortion than ever before.

We show how transformers can be used to vastly simplify neural video compression. Previous methods have been relying on an increasing number of architectural biases and priors, including motion prediction and warping operations, resulting in complex models. Instead, we independently map input frames to representations and use a transformer to model their dependencies, letting it predict the distribution of future representations given the past. The resulting video compression transformer outperforms previous methods on standard video compression data sets. Experiments on synthetic data show that our model learns to handle complex motion patterns such as panning, blurring and fading purely from data. Our approach is easy to implement, and we release code to facilitate future research.

Despite considerable progress on end-to-end optimized deep networks for image compression, video coding remains a challenging task. Recently proposed methods for learned video compression use optical flow and bilinear warping for motion compensation and show competitive rate-distortion performance relative to hand-engineered codecs like H.264 and HEVC. However, these learning-based methods rely on complex architectures and training schemes including the use of pre-trained optical flow networks, sequential training of sub-networks, adaptive rate control, and buffering intermediate reconstructions to disk during training. In this paper, we show that a generalized warping operator that better handles common failure cases, e.g. disocclusions and fast motion, can provide competitive compression results with a greatly simplified model and training procedure. Specifically, we propose scale-space flow, an intuitive generalization of optical flow that adds a scale parameter to allow the network to better model uncertainty. Our experiments show that a low-latency video compression model (no B-frames) using scale-space flow for motion compensation can outperform analogous state-of-the art learned video compression models while being trained using a much simpler procedure and without any pre-trained optical flow networks.

Pre-trained convolutional neural networks (CNNs) are powerful off-the-shelf feature generators and have been shown to perform very well on a variety of tasks. Unfortunately, the generated features are high dimensional and expensive to store: potentially hundreds of thousands of floats per example when processing videos. Traditional entropy based lossless compression methods are of little help as they do not yield desired level of compression, while general purpose lossy compression methods based on energy compaction (e.g. PCA followed by quantization and entropy coding) are sub-optimal, as they are not tuned to task specific objective. We propose a learned method that jointly optimizes for compressibility along with the task objective for learning the features. The plug-in nature of our method makes it straight-forward to integrate with any target objective and trade-off against compressibility. We present results on multiple benchmarks and demonstrate that our method produces features that are an order of magnitude more compressible, while having a regularization effect that leads to a consistent improvement in accuracy.

We extensively study how to combine Generative Adversarial Networks and learned compression to obtain a state-of-the-art generative lossy compression system. In particular, we investigate normalization layers, generator and discriminator architectures, training strategies, as well as perceptual losses. In contrast to previous work, i) we obtain visually pleasing reconstructions that are perceptually similar to the input, ii) we operate in a broad range of bitrates, and iii) our approach can be applied to high-resolution images. We bridge the gap between rate-distortion-perception theory and practice by evaluating our approach both quantitatively with various perceptual metrics, and with a user study. The study shows that our method is preferred to previous approaches even if they use more than 2x the bitrate.

This paper introduces a video dataset of spatio-temporally localized Atomic Visual Actions (AVA). The AVA dataset densely annotates 80 atomic visual actions in 437 15-minute video clips, where actions are localized in space and time, resulting in 1.59M action labels with multiple labels per person occurring frequently. The key characteristics of our dataset are: (1) the definition of atomic visual actions, rather than composite actions; (2) precise spatio-temporal annotations with possibly multiple annotations for each person; (3) exhaustive annotation of these atomic actions over 15-minute video clips; (4) people temporally linked across consecutive segments; and (5) using movies to gather a varied set of action representations. This departs from existing datasets for spatio-temporal action recognition, which typically provide sparse annotations for composite actions in short video clips. AVA, with its realistic scene and action complexity, exposes the intrinsic difficulty of action recognition. To benchmark this, we present a novel approach for action localization that builds upon the current state-of-the-art methods, and demonstrates better performance on JHMDB and UCF101-24 categories. While setting a new state of the art on existing datasets, the overall results on AVA are low at 15.8% mAP, underscoring the need for developing new approaches for video understanding.

We propose a method for lossy image compression based on recurrent, convolutional neural networks that outperforms BPG (4:2:0), WebP, JPEG2000, and JPEG as measured by MS-SSIM. We introduce three improvements over previous research that lead to this state-of-the-art result using a single model. First, we modify the recurrent architecture to improve spatial diffusion, which allows the network to more effectively capture and propagate image information through the network's hidden state. Second, in addition to lossless entropy coding, we use a spatially adaptive bit allocation algorithm to more efficiently use the limited number of bits to encode visually complex image regions. Finally, we show that training with a pixel-wise loss weighted by SSIM increases reconstruction quality according to multiple metrics. We evaluate our method on the Kodak and Tecnick image sets and compare against standard codecs as well as recently published methods based on deep neural networks.

We present a full reference, perceptual image metric based on VGG-16, an artificial neural network trained on object classification. We fit the metric to a new database based on 140k unique images annotated with ground truth by human raters who received minimal instruction. The resulting metric shows competitive performance on TID 2013, a database widely used to assess image quality assessments methods. More interestingly, it shows strong responses to objects potentially carrying semantic relevance such as faces and text, which we demonstrate using a visualization technique and ablation experiments. In effect, the metric appears to model a higher influence of semantic context on judgments, which we observe particularly in untrained raters. As the vast majority of users of image processing systems are unfamiliar with Image Quality Assessment (IQA) tasks, these findings may have significant impact on real-world applications of perceptual metrics.

The leading approach for image compression with artificial neural networks (ANNs) is to learn a nonlinear transform and a fixed entropy model that are optimized for rate-distortion performance. We show that this approach can be significantly improved by incorporating spatially local, image-dependent entropy models. The key insight is that existing ANN-based methods learn an entropy model that is shared between the encoder and decoder, but they do not transmit any side information that would allow the model to adapt to the structure of a specific image. We present a method for augmenting ANN-based image coders with image-dependent side information that leads to a 17.8% rate reduction over a state-of-the-art ANN-based baseline model on a standard evaluation set, and 70-98% reductions on images with low visual complexity that are poorly captured by a fixed, global entropy model.

Recent models for learned image compression are based on autoencoders that learn approximately invertible mappings from pixels to a quantized latent representation. The transforms are combined with an entropy model, which is a prior on the latent representation that can be used with standard arithmetic coding algorithms to generate a compressed bitstream. Recently, hierarchical entropy models were introduced as a way to exploit more structure in the latents than previous fully factorized priors, improving compression performance while maintaining end-to-end optimization. Inspired by the success of autoregressive priors in probabilistic generative models, we examine autoregressive, hierarchical, and combined priors as alternatives, weighing their costs and benefits in the context of image compression. While it is well known that autoregressive models can incur a significant computational penalty, we find that in terms of compression performance, autoregressive and hierarchical priors are complementary and can be combined to exploit the probabilistic structure in the latents better than all previous learned models. The combined model yields state-of-the-art rate-distortion performance and generates smaller files than existing methods: 15.8% rate reductions over the baseline hierarchical model and 59.8%, 35%, and 8.4% savings over JPEG, JPEG2000, and BPG, respectively. To the best of our knowledge, our model is the first learning-based method to outperform the top standard image codec (BPG) on both the PSNR and MS-SSIM distortion metrics.

This paper presents a set of full-resolution lossy image compression methods based on neural networks. Each of the architectures we describe can provide variable compression rates during deployment without requiring retraining of the network: each network need only be trained once. All of our architectures consist of a recurrent neural network (RNN)-based encoder and decoder, a binarizer, and a neural network for entropy coding. We compare RNN types (LSTM, associative LSTM) and introduce a new hybrid of GRU and ResNet. We also study “one-shot” versus additive reconstruction architectures and introduce a new scaled-additive framework. We compare to previous work, showing improvements of 4.3%-8.8% AUC (area under the rate-distortion curve), depending on the perceptual metric used. As far as we know, this is the first neural network architecture that is able to outperform JPEG at image compression across most bitrates on the rate-distortion curve on the Kodak dataset images, with and without the aid of entropy coding.

Deep neural networks represent a powerful class of function approximators that can learn to compress and reconstruct images. Existing image compression algorithms based on neural networks learn quantized representations with a constant spatial bit rate across each image. While entropy coding introduces some spatial variation, traditional codecs have benefited significantly by explicitly adapting the bit rate based on local image complexity and visual saliency. This paper introduces an algorithm that combines deep neural networks with quality-sensitive bit rate adaptation using a tiled network. We demonstrate the importance of spatial context prediction and show improved quantitative (PSNR) and qualitative (subjective rater assessment) results compared to a non-adaptive baseline and a recently published image compression model based on fully-convolutional neural networks.

Convolutional neural networks (CNNs) have been extensively applied for image recognition problems giving state-of-the-art results on recognition, detection, segmentation and retrieval. In this work we propose and evaluate several deep neural network architectures to combine image information across a video over longer time periods than previously attempted. We propose two methods capable of handling full length videos. The first method explores various convolutional temporal feature pooling architectures, examining the various design choices which need to be made when adapting a CNN for this task. The second proposed method explicitly models the video as an ordered sequence of frames. For this purpose we employ a recurrent neural network that uses Long Short-Term Memory (LSTM) cells which are connected to the output of the underlying CNN. Our best networks exhibit significant performance improvements over previously published results on the Sports 1 million dataset (73.1% vs. 60.9%) and the UCF-101 datasets with (88.6% vs. 88.0%) and without additional optical flow information (82.6% vs. 73.0%).

No abstract available.

Convolutional Neural Networks (CNNs) have been established as a powerful class of models for image recognition problems. Encouraged by these results, we provide an extensive empirical evaluation of CNNs on large-scale video classification using a new dataset of 1 million YouTube videos belonging to 487 classes. We study multiple approaches for extending the connectivity of a CNN in time domain to take advantage of local spatio-temporal information and suggest a multiresolution, foveated architecture as a promising way of speeding up the training. Our best spatio-temporal networks display significant performance improvements compared to strong feature-based baselines (55.3% to 63.9%), but only a surprisingly modest improvement compared to single-frame models (59.3% to 60.9%). We further study the generalization performance of our best model by retraining the top layers on the UCF-101 Action Recognition dataset and observe significant performance improvements compared to the UCF-101 baseline model (63.3% up from 43.9%).

We consider the problem of content-based automated tag learning. In particular, we address semantic variations (sub-tags) of the tag. Each video in the training set is assumed to be associated with a sub-tag label, and we treat this sub-tag label as latent information. A latent learning framework based on LogitBoost is proposed, which jointly considers both the tag label and the latent sub-tag label. The latent sub-tag information is exploited in our framework to assist the learning of our end goal, i.e., tag prediction. We use the cowatch information to initialize the learning process. In experiments, we show that the proposed method achieves significantly better results over baselines on a large-scale testing video set which contains about 50 million YouTube videos.

We present a system that automatically recommends tags for YouTube videos solely based on their audiovisual content. We also propose a novel framework for unsupervised discovery of video categories that exploits knowledge mined from the World-Wide Web text documents/searches. First, video content to tag association is learned by training classifiers that map audiovisual content-based features from millions of videos on YouTube.com to existing uploader-supplied tags for these videos. When a new video is uploaded, the labels provided by these classifiers are used to automatically suggest tags deemed relevant to the video. Our system has learned a vocabulary of over 20,000 tags. Secondly, we mined large volumes of Web pages and search queries to discover a set of possible text entity categories and a set of associated is-A relationships that map individual text entities to categories. Finally, we apply these is-A relationships mined from web text on the tags learned from audiovisual content of videos to automatically synthesize a reliable set of categories most relevant to videos – along with a mechanism to predict these categories for new uploads. We then present rigorous rating studies that establish that: (a) the average relevance of tags automatically recommended by our system matches the average relevance of the uploader-supplied tags at the same or better coverage and (b) the average precision@K of video categories discovered by our system is 70% with K=5.

In this paper, we present a new method for bidirectional relighting for 3D-aided 2D face recognition under large pose and illumination changes. During subject enrollment, we build subject-specific 3D annotated models by using the subjects' raw 3D data and 2D texture. During authentication, the probe 2D images are projected onto a normalized image space using the subject-specific 3D model in the gallery. Then, a bidirectional relighting algorithm and two similarity metrics (a view-dependent complex wavelet structural similarity and a global similarity) are employed to compare the gallery and probe. We tested our algorithms on the UHDB11 and UHDB12 databases that contain 3D data with probe images under large lighting and pose variations. The experimental results show the robustness of our approach in recognizing faces in difficult situations.

This paper presents an efficient, personalizable and yet completely automatic algorithm for enhancing the brightness, tonal balance, and contrast of faces in thumbnails of online videos where multiple colored illumination sources are the norm and artifacts such as poor illumination and backlight are common. These artifacts significantly lower the perceptual quality of faces and skin, and cannot be easily corrected by common global image transforms. The same identifiable user, however, often uploads or participates in multiple photos, videos, or video chat sessions with varying illumination conditions. The proposed algorithm adaptively transforms the skin pixels in a poor illumination environment to match the skin color model of a prototypical face of the same user in a better illumination environment. It leaves the remaining non-skin portions of the image virtually unchanged while ascertaining a smooth, natural appearance. A component of our system automatically selects such a prototypical face for each user given a collection of uploaded videos/photo albums or prior video chat sessions by that user. We present several human rating studies on YouTube data that quantitatively demonstrate significant improvement in facial quality using the proposed algorithm.

A fast and robust method for video contrast enhancement is presented. The method uses the histogram of each frame, along with upper and lower bounds computed per shot in order to enhance the current frame. This ensures that the artifacts introduced during the enhancement is reduced to a minimum. Traditional methods that do not compute per-shot estimates tend to over-enhance parts of the video such as fades and transitions. Our method does not suffer from this problem, which is essential for a fully automatic algorithm. We present the parameters for our methods which yielded the best human feedback, which showed that out of 208 videos, 203 were enhanced, while the remaining 5 were of too poor quality to be enhanced. Additionally, we present a visual comparison of our work with the recently-proposed Weighted Thresholded Histogram Equalization (WTHE) algorithm.

Evaluation of face recognition performance under varying illumination, poses, and image degradation due to sensor characteristics has been studied in depth for static images. This paper uses static and video images when evaluating a face recognition algorithm based on ¿Adaptive Principal Components Analysis¿ (APCA). The results show that the same challenges that happen in static images apply to video, and that when multiple images are available the algorithms perform better (e.g. using, mean frames for face recognition).

Three-dimensional data are increasingly being used for biometric purposes as they offer resilience to problems common mon in two-dimensional data. They have been successfully applied to face recognition and more recently to ear recognition. However, real-life biometric applications require algorithms that are both robust and efficient so that they scale well with the size of the databases. A novel ear recognition method is presented that uses a generic annotated ear model to register and fit each ear dataset. Then a compact biometric signature is extracted that retains 3D information. The proposed method is evaluated using the largest publicly available 3D ear database appended with our own database, resulting in a database containing data from multiple 3D sensor types. Using this database it is shown that the proposed method is not only robust, accurate and sensor invariant but also extremely efficient, thus making it suitable for real-life biometric applications.

In this paper, we present the computational tools and a hardware prototype for 3D face recognition. Full automation is provided through the use of advanced multistage alignment algorithms, resilience to facial expressions by employing a deformable model framework, and invariance to 3D capture devices through suitable preprocessing steps. In addition, scalability in both time and space is achieved by converting 3D facial scans into compact metadata. We present our results on the largest known, and now publicly available, Face Recognition Grand Challenge 3D facial database consisting of several thousand scans. To the best of our knowledge, this is the highest performance reported on the FRGC v2 database for the 3D modality.

In this paper, we present a new 3D face recognition approach. Full automation is provided through the use of advanced multi-stage alignment algorithms, resilience to facial expressions by employing a deformable model framework, and invariance to 3D capture devices through suitable preprocessing steps. In addition, scalability in both time and space is achieved by converting 3D facial scans into compact wavelet metadata. We present results on the largest known, and now publicly-available, Face Recognition Grand Challenge 3D facial database consisting of several thousand scans. To the best of our knowledge, our approach has achieved the highest accuracy on this dataset. 1

It is becoming increasingly important to be able to credential and identify authorized personnel at key points of entry. Such identity management systems commonly employ biometric identifiers. In this paper, we present a novel multimodal facial recognition approach that employs data from both visible spectrum and thermal infrared sensors. Data from multiple cameras is used to construct a three-dimensional mesh representing the face and a facial thermal texture map. An annotated face model with explicit two-dimensional parameterization (UV) is then fitted to this data to construct: 1) a three-channel UV deformation image encoding geometry, and 2) a one-channel UV vasculature image encoding facial vasculature. Recognition is accomplished by comparing: 1) the parametric deformation images, 2) the parametric vasculature images, and 3) the visible spectrum texture maps. The novelty of our work lies in the use of deformation images and physiological information as means for comparison. We have performed extensive tests on the Face Recognition Grand Challenge v1.0 dataset and on our own multimodal database with very encouraging results.

Biometrics-based technologies in the area of identity management are gaining increasing importance, as a means of establishing non-falsifiable credentials for end users. However, in the three-way tug-of-war between convenient, unobtrusive data collection (required for user acceptance), accuracy in results (required for justifying deployment), and speed (required for widespread use in practice), no single biometric to date has managed to hold the middle ground that would allow for its ready adoption. The overall goal of our project is to develop the theoretical framework and computational tools that will lead to the development of a practical, unobtrusive, and accurate face recognition system for convenient and effective access control. This framework encompasses 8D characteristics of the face (3D geometry+2D visible texture+2D infrared texture, over time). In this paper, we present a novel multi-modal facial recognition approach that employs data from both visible spectrum and thermal infrared sensors. From the fitted parametric model we extract two images corresponding to the subject's face and process these images to extract biometric signatures. Specifically, the deformation image is compressed using a wavelet transform and the vasculature graph is extracted from the parametric thermal image.