CVPR 2016 Papers

Long training times for high-accuracy deep neural networks (DNNs) impede research into new DNN architectures and slow the development of high-accuracy DNNs. In this paper we present FireCaffe, which successfully scales deep neural network training across a cluster of GPUs. We also present a number of best practices to aid in comparing advancements in methods for scaling and accelerating the training of deep neural networks. The speed and scalability of distributed algorithms is almost always limited by the overhead of communicating between servers; DNN training is not an exception to this rule. Therefore, the key consideration here is to reduce communication overhead wherever possible, while not degrading the accuracy of the DNN models that we train. Our approach has three key pillars. First, we select network hardware that achieves high bandwidth between GPU servers -- Infiniband or Cray interconnects are ideal for this. Second, we consider a number of communication algorithms, and we find that reduction trees are more efficient and scalable than the traditional parameter server approach. Third, we optionally increase the batch size to reduce the total quantity of communication during DNN training, and we identify hyperparameters that allow us to reproduce the small-batch accuracy while training with large batch sizes. When training GoogLeNet and Network-in-Network on ImageNet, we achieve a 47x and 39x speedup, respectively, when training on a cluster of 128 GPUs.

Given a food image, can a fine-grained object recognition engine tell "which restaurant which dish" the food belongs to? Such ultra-fine grained image recognition is the key for many applications like search by images, but it is very challenging because it needs to discern subtle difference between classes while dealing with the scarcity of training data. Fortunately, the ultra-fine granularity naturally brings rich relationships among object classes. This paper proposes a novel approach to exploit the rich relationships through bipartite-graph labels (BGL). We show how to model BGL in an overall convolutional neural networks and the resulting system can be optimized through back-propagation. We also show that it is computationally efficient in inference thanks to the bipartite structure. To facilitate the study, we construct a new food benchmark dataset, which consists of 37,885 food images collected from 6 restaurants and totally 975 menus. Experimental results on this new food and three other datasets demonstrate BGL advances previous works in fine-grained object recognition. An online demo is available at http://www.f-zhou.com/fg_demo/.

Existing fine-grained visual categorization methods often suffer from three challenges: lack of training data, large number of fine-grained categories, and high intra-class vs. low inter-class variance. In this work we propose a generic iterative framework for fine-grained categorization and dataset bootstrapping that handles these three challenges. Using deep metric learning with humans in the loop, we learn a low dimensional feature embedding with anchor points on manifolds for each category. These anchor points capture intra-class variances and remain discriminative between classes. In each round, images with high confidence scores from our model are sent to humans for labeling. By comparing with exemplar images, labelers mark each candidate image as either a "true positive" or a "false positive." True positives are added into our current dataset and false positives are regarded as "hard negatives" for our metric learning model. Then the model is re-trained with an expanded dataset and hard negatives for the next round. To demonstrate the effectiveness of the proposed framework, we bootstrap a fine-grained flower dataset with 620 categories from Instagram images. The proposed deep metric learning scheme is evaluated on both our dataset and the CUB-200-2001 Birds dataset. Experimental evaluations show significant performance gain using dataset bootstrapping and demonstrate state-of-the-art results achieved by the proposed deep metric learning methods.

In many large-scale video analysis scenarios, one is interested in localizing and recognizing human activities that occur in short temporal intervals within long untrimmed videos. Current approaches for activity detection still struggle to handle large-scale video collections and the task remains relatively unexplored. This is in part due to the computational complexity of current action recognition approaches and the lack of a method that proposes fewer intervals in the video, where activity processing can be focused. In this paper, we introduce a proposal method that aims to recover temporal segments containing actions in untrimmed videos. Building on techniques for learning sparse dictionaries, we introduce a learning framework to represent and retrieve activity proposals. We demonstrate the capabilities of our method in not only producing high quality proposals but also in its efficiency. Finally, we show the positive impact our method has on recognition performance when it is used for action detection, while running at 10FPS.

Detecting edges is a fundamental problem in computer vision with many applications, some involving very noisy images. While most edge detection methods are fast, they perform well only on relatively clean images. Unfortunately, sophisticated methods that are robust to high levels of noise are quite slow. In this paper we develop a novel multiscale method to detect curved edges in noisy images. Even though our algorithm searches for edges over an exponentially large set of candidate curves, its runtime is nearly linear in the total number of image pixels. As we demonstrate experimentally, our algorithm is orders of magnitude faster than previous methods designed to deal with high noise levels. At the same time it obtains comparable and often superior results to existing methods on a variety of challenging noisy images.

We revisit the idea of brain damage, i.e. the pruning of the coefficients of a neural network, and suggest how brain damage can be modified and used to speedup convolutional layers in ConvNets. The approach uses the fact that many efficient implementations reduce generalized convolutions to matrix multiplications. The suggested brain damage process prunes the convolutional kernel tensor in a group-wise fashion. After such pruning, convolutions can be reduced to multiplications of thinned dense matrices, which leads to speedup. We investigate different ways to add group-wise prunning to the learning process, and show that several-fold speedups of convolutional layers can be attained using group-sparsity regularizers. Our approach can adjust the shapes of the receptive fields in the convolutional layers, and even prune excessive feature maps from ConvNets, all in a data-driven way.

The SVM+ approach has shown excellent performance in visual recognition tasks for exploiting privileged information in the training data. In this paper, we propose two efficient algorithms for solving the linear and kernel SVM+, respectively. For linear SVM+, we absorb the bias term into the weight vector, and formulate a new optimization problem with simpler constraints in the dual form. Then, we develop an efficient dual coordinate descent algorithm to solve the new optimization problem. For kernel SVM+, we further apply the l2-loss, which leads to a simpler optimization problem in the dual form with only half of dual variables when compared with the dual form of the original SVM+ method. More interestingly, we show that our new dual problem can be efficiently solved by using the SMO algorithm of the one-class SVM problem. Comprehensive experiments on three datasets clearly demonstrate that our proposed algorithms achieve significant speed-up than the state-of-the-art solvers for linear and kernel SVM+.

We propose a novel approach for learning features from weakly-supervised data by joint ranking and classification. In order to exploit data with weak labels, we jointly train a feature extraction network with a ranking loss and a classification network with a cross-entropy loss. We obtain high-quality compact discriminative features with few parameters, learned on relatively small datasets without additional annotations. This enables us to tackle tasks with specialized images not very similar to the more generic ones in existing fully-supervised datasets. We show that the resulting features in combination with a linear classifier surpass the state-of-the-art on the Hipster Wars dataset despite using features only 0.3% of the size. Our proposed features significantly outperform those obtained from networks trained on ImageNet, despite being 32 times smaller (128 single-precision floats), trained on noisy and weakly-labeled data, and using only 1.5% of the number of parameters.

Previous studies on facial expression analysis have been focused on recognizing basic expression categories. There is limited amount of work on the continuous expression intensity estimation, which is important for detecting and tracking emotion change. Part of the reason is the lack of labeled data with annotated expression intensity since expression intensity annotation requires expertise and is time consuming. In this work, we treat the expression intensity estimation as a regression problem. By taking advantage of the natural onset-apex-offset evolution pattern of facial expression, the proposed method can handle different amounts of annotations to perform frame-level expression intensity estimation. In fully supervised case, all the frames are provided with intensity annotations. In weakly supervised case, only the annotations of selected key frames are used. While in unsupervised case, expression intensity can be estimated without any annotations. An efficient optimization algorithm based on Alternating Direction Method of Multipliers (ADMM) is developed for solving the optimization problem associated with parameter learning. We demonstrate the effectiveness of proposed method by comparing it against both fully supervised and unsupervised approaches on benchmark facial expression datasets.

From scientific research to commercial applications, eye tracking is an important tool across many domains. Despite its range of applications, eye tracking has yet to become a pervasive technology. We believe that we can put the power of eye tracking in everyone's palm by building eye tracking software that works on commodity hardware such as mobile phones and tablets, without the need for additional sensors or devices. We tackle this problem by introducing GazeCapture, the first large-scale dataset for eye tracking, containing data from over 1450 people consisting of almost 2:5M frames. Using GazeCapture, we train iTracker, a convolutional neural network for eye tracking, which achieves a significant reduction in error over previous approaches while running in real time (10-15fps) on a modern mobile device. Our model achieves a prediction error of 1.71cm and 2.53cm without calibration on mobile phones and tablets respectively. With calibration, this is reduced to 1.34cm and 2.12cm. Further, we demonstrate that the features learned by iTracker generalize well to other datasets, achieving state-of-the-art results. The code, data, and models are available at http://gazecapture.csail.mit.edu.

Conventional scanning and multiplexing techniques for hyperspectral imaging suffer from limited temporal and/or spatial resolution. To resolve this issue, coding techniques are becoming increasingly popular in developing snapshot systems for high-resolution hyperspectral imaging. For such systems, it is a critical task to accurately restore the 3D hyperspectral image from its corresponding coded 2D image. In this paper, we propose an effective method for coded hyperspectral image restoration, which exploits extensive structure sparsity in the hyperspectral image. Specifically, we simultaneously explore spectral and spatial correlation via low-rank regularizations, and formulate the restoration problem into a variational optimization model, which can be solved via an iterative numerical algorithm. Experimental results using both synthetic data and real images show that the proposed method can significantly outperform the state-of-the-art methods on several popular coding-based hyperspectral imaging systems.

Convolutional neural networks (CNNs) have shown great performance as general feature representations for object recognition applications. However, for multi-label images that contain multiple objects from different categories, scales and locations, global CNN features are not optimal. In this paper, we incorporate local information to enhance the feature discriminative power. In particular, we first extract object proposals from each image. With each image treated as a bag and object proposals extracted from it treated as instances, we transform the multi-label recognition problem into a multi-class multi-instance learning problem. Then, in addition to extracting the typical CNN feature representation from each proposal, we propose to make use of ground-truth bounding box annotations (strong labels) to add another level of local information by using nearest-neighbor relationships of local regions to form a multi-view pipeline. The proposed multi-view multi-instance framework utilizes both weak and strong labels effectively, and more importantly it has the generalization ability to even boost the performance of unseen categories by partial strong labels from other categories. Our framework is extensively compared with state-of-the-art hand-crafted feature based methods and CNN based methods on two multi-label benchmark datasets. The experimental results validate the discriminative power and the generalization ability of the proposed framework. With strong labels, our framework is able to achieve state-of-the-art results in both datasets.

In this paper, we investigate two new strategies to detect objects accurately and efficiently using deep convolutional neural network: 1) scale-dependent pooling and 2) layer-wise cascaded rejection classifiers. The scale-dependent pooling (SDP) improves detection accuracy by exploiting appropriate convolutional features depending on the scale of candidate object proposals. The cascaded rejection classifiers (CRC) effectively utilize convolutional features and eliminate negative object proposals in a cascaded manner, which greatly speeds up the detection while maintaining high accuracy. In combination of the two, our method achieves significantly better accuracy compared to other state-of-the-arts in three challenging datasets, PASCAL object detection challenge, KITTI object detection benchmark and newly collected Inner-city dataset, while being more efficient.

When creating a photo album of an event, people typically select a few important images to keep or share. There is some consistency in the process of choosing the important images, and discarding the unimportant ones. Modeling this selection process will assist automatic photo selection and album summarization. In this paper, we show that the selection of important images is consistent among different viewers, and that this selection process is related to the event type of the album. We introduce the concept of event-specific image importance. We collected a new event album dataset with human annotation of the relative image importance with each event album. We also propose a Convolutional Neural Network (CNN) based method to predict the image importance score of a given event album, using a novel rank loss function and a progressive training scheme. Results demonstrate that our method significantly outperforms various baseline methods.

Conventional algorithms for sparse signal recovery and sparse representation rely on l1-norm regularized variational methods. However, when applied to the reconstruction of sparse images, i.e., images where only a few pixels are non-zero, simple l1-norm-based methods ignore poten- tial correlations in the support between adjacent pixels. In a number of applications, one is interested in images that are not only sparse, but also have a support with smooth (or contiguous) boundaries. Existing algorithms that take into account such a support structure mostly rely on non-convex methods and--as a consequence--do not scale well to high-dimensional problems and/or do not converge to global optima. In this paper, we explore the use of new block l1-norm regularizers, which enforce image sparsity while simultaneously promoting smooth support structure. By exploiting the convexity of our regularizers, we develop new computationally-efficient recovery algorithms that guarantee global optimality. We demonstrate the efficacy of our regularizers on a variety of imaging tasks including compressive image recovery, image restoration, and robust PCA.

During the past few years we have witnessed the development of many methodologies for building and fitting Statistical Deformable Models (SDMs). The construction of accurate SDMs requires careful annotation of images with regards to a consistent set of landmarks. However, the manual annotation of a large amount of images is a tedious, laborious and expensive procedure. Furthermore, for several deformable objects, e.g. human body, it is difficult to define a consistent set of landmarks, and, thus, it becomes impossible to train humans in order to accurately annotate a collection of images. Nevertheless, for the majority of objects, it is possible to extract the shape by object segmentation or even by shape drawing. In this paper, we show for the first time, to the best of our knowledge, that it is possible to construct SDMs by putting object shapes in dense correspondence. Such SDMs can be built with much less effort for a large battery of objects. Additionally, we show that, by sampling the dense model, a part-based SDM can be learned with its parts being in correspondence. We employ our framework to develop SDMs of human arms and legs, which can be used for the segmentation of the outline of the human body, as well as to provide better and more consistent annotations for body joints.

Most existing dictionary learning algorithms consider a linear sparse model, which often cannot effectively characterize the nonlinear properties present in many types of visual data, e.g. dynamic texture (DT). Such nonlinear properties can be exploited by the so-called kernel sparse coding. This paper proposed an equiangular kernel dictionary learning method with optimal mutual coherence to exploit the nonlinear sparsity of high-dimensional visual data. Two main issues are addressed in the proposed method: (1) coding stability for redundant dictionary of infinite-dimensional space; and (2) computational efficiency for computing kernel matrix of training samples of high-dimensional data. The proposed kernel sparse coding method is applied to dynamic texture analysis with both local DT pattern extraction and global DT pattern characterization. The experimental results showed its performance gain over existing methods.

In this work we introduce a fully end-to-end approach for action detection in videos that learns to directly predict the temporal bounds of actions. Our intuition is that the process of detecting actions is naturally one of observation and refinement: observing moments in video, and refining hypotheses about when an action is occurring. Based on this insight, we formulate our model as a recurrent neural network-based agent that interacts with a video over time. The agent observes video frames and decides both where to look next and whether to emit a prediction. Since backpropagation is not adequate in this non-differentiable setting, we use REINFORCE to learn the agent's task-specific decision policy. Our model achieves state-of-the-art results on the THUMOS'14 and ActivityNet datasets while observing only a fraction (2% or less) of the video frames.

Recent algorithms in convolutional neural networks (CNN) considerably advance the fine-grained image classification, which aims to differentiate the subtle differences among subordinate classes. However, previous studies have rarely focused on learning a fined-grained and structured feature representation that is able to locate relevant images at different levels of relevance, e.g., discovering cars from the same make or the same model, both of which require high precision. In this paper, we propose two main contributions to tackle this problem. 1) A multi-task learning framework is designed to effectively learn fine-grained feature representations by jointly optimizing both classification and similarity constraints. 2) To model the multi-level relevance, label structures such as hierarchy or shared attributes are seamlessly embedded into the framework by generalizing the triplet loss. Extensive and thorough experiments have been conducted on three fine-grained datasets, i.e., the Stanford car, the car-333, and the food datasets, which contain either hierarchical labels or shared attributes. Our proposed method has achieved very competitive performance, i.e., among state-of-the-art classification accuracy. More importantly, it significantly outperforms previous fine-grained feature representations for image retrieval at different levels of relevance