While much of the work in the design of convolutional networks over the last five years has revolved around the empirical investigation of the importance of depth, filter sizes, and number of feature channels, recent studies have shown that branching, i.e., splitting the computation along parallel but distinct threads and then aggregating their outputs, represents a new promising dimension for significant improvements in performance. To combat the complexity of design choices in multi-branch architectures, prior work has adopted simple strategies, such as a fixed branching factor, the same input being fed to all parallel branches, and an additive combination of the outputs produced by all branches at aggregation points. In this work we remove these predefined choices and propose an algorithm to learn the connections between branches in the network. Instead of being chosen a priori by the human designer, the multi-branch connectivity is learned simultaneously with the weights of the network by optimizing a single loss function defined with respect to the end task. We demonstrate our approach on the problem of multi-class image classification using three different datasets where it yields consistently higher accuracy compared to the state-of-the-art "ResNeXt" multi-branch network given the same learning capacity.
We study the design of deep architectures for lossy image compression. We present two architectural recipes in the context of multi-stage progressive encoders and empirically demonstrate their importance on compression performance. Specifically, we show that: (a) predicting the original image data from residuals in a multi-stage progressive architecture facilitates learning and leads to improved performance at approximating the original content and (b) learning to inpaint (from neighboring image pixels) before performing compression reduces the amount of information that must be stored to achieve a high-quality approximation. Incorporating these design choices in a baseline progressive encoder yields an average reduction of over $60\%$ in file size with similar quality compared to the original residual encoder.
While there is overall agreement that future technology for organizing, browsing and searching videos hinges on the development of methods for high-level semantic understanding of video, so far no consensus has been reached on the best way to train and assess models for this task. Casting video understanding as a form of action or event categorization is problematic as it is not fully clear what the semantic classes or abstractions in this domain should be. Language has been exploited to sidestep the problem of defining video categories, by formulating video understanding as the task of captioning or description. However, language is highly complex, redundant and sometimes ambiguous. Many different captions may express the same semantic concept. To account for this ambiguity, quantitative evaluation of video description requires sophisticated metrics, whose performance scores are typically hard to interpret by humans. This paper provides four contributions to this problem. First, we formulate Video Multiple Choice Caption (VideoMCC) as a new well-defined task with an easy-to-interpret performance measure. Second, we describe a general semi-automatic procedure to create benchmarks for this task. Third, we publicly release a large-scale video benchmark created with an implementation of this procedure and we include a human study that assesses human performance on our dataset. Finally, we propose and test a varied collection of approaches on this benchmark for the purpose of gaining a better understanding of the new challenges posed by video comprehension.
Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.
We present a tree-structured network architecture for large scale image classification. The trunk of the network contains convolutional layers optimized over all classes. At a given depth, the trunk splits into separate branches, each dedicated to discriminate a different subset of classes. Each branch acts as an expert classifying a set of categories that are difficult to tell apart, while the trunk provides common knowledge to all experts in the form of shared features. The training of our "network of experts" is completely end-to-end: the partition of categories into disjoint subsets is learned simultaneously with the parameters of the network trunk and the experts are trained jointly by minimizing a single learning objective over all classes. The proposed structure can be built from any existing convolutional neural network (CNN). We demonstrate its generality by adapting 4 popular CNNs for image categorization into the form of networks of experts. Our experiments on CIFAR100 and ImageNet show that in every case our method yields a substantial improvement in accuracy over the base CNN, and gives the best result achieved so far on CIFAR100. Finally, the improvement in accuracy comes at little additional cost: compared to the base network, the training time is only moderately increased and the number of parameters is comparable or in some cases even lower.
Histopathological characterization of colorectal polyps is an important principle for determining the risk of colorectal cancer and future rates of surveillance for patients. This characterization is time-intensive, requires years of specialized training, and suffers from significant inter-observer and intra-observer variability. In this work, we built an automatic image-understanding method that can accurately classify different types of colorectal polyps in whole-slide histology images to help pathologists with histopathological characterization and diagnosis of colorectal polyps. The proposed image-understanding method is based on deep-learning techniques, which rely on numerous levels of abstraction for data representation and have shown state-of-the-art results for various image analysis tasks. Our image-understanding method covers all five polyp types (hyperplastic polyp, sessile serrated polyp, traditional serrated adenoma, tubular adenoma, and tubulovillous/villous adenoma) that are included in the US multi-society task force guidelines for colorectal cancer risk assessment and surveillance, and encompasses the most common occurrences of colorectal polyps. Our evaluation on 239 independent test samples shows our proposed method can identify the types of colorectal polyps in whole-slide images with a high efficacy (accuracy: 93.0%, precision: 89.7%, recall: 88.3%, F1 score: 88.8%). The presented method in this paper can reduce the cognitive burden on pathologists and improve their accuracy and efficiency in histopathological characterization of colorectal polyps, and in subsequent risk assessment and follow-up recommendations.
In this work we focus on the problem of colorization for image compression. Since color information occupies a large proportion of the total storage size of an image, a method that can predict accurate color from its grayscale version can produce dramatic reduction in image file size. But colorization for compression poses several challenges. First, while colorization for artistic purposes simply involves predicting plausible chroma, colorization for compression requires generating output colors that are as close as possible to the ground truth. Second, many objects in the real world exhibit multiple possible colors. Thus, to disambiguate the colorization problem some additional information must be stored to reproduce the true colors with good accuracy. To account for the multimodal color distribution of objects we propose a deep tree-structured network that generates multiple color hypotheses for every pixel from a grayscale picture (as opposed to a single color produced by most prior colorization approaches). We show how to leverage the multimodal output of our model to reproduce with high fidelity the true colors of an image by storing very little additional information. In the experiments we show that our proposed method outperforms traditional JPEG color coding by a large margin, producing colors that are nearly indistinguishable from the ground truth at the storage cost of just a few hundred bytes for high-resolution pictures!
In many computer vision tasks, the relevant information to solve the problem at hand is mixed to irrelevant, distracting information. This has motivated researchers to design attentional models that can dynamically focus on parts of images or videos that are salient, e.g., by down-weighting irrelevant pixels. In this work, we propose a spatiotemporal attentional model that learns where to look in a video directly from human fixation data. We model visual attention with a mixture of Gaussians at each frame. This distribution is used to express the probability of saliency for each pixel. Time consistency in videos is modeled hierarchically by: 1) deep 3D convolutional features to represent spatial and short-term time relations and 2) a long short-term memory network on top that aggregates the clip-level representation of sequential clips and therefore expands the temporal domain from few frames to seconds. The parameters of the proposed model are optimized via maximum likelihood estimation using human fixations as training data, without knowledge of the action in each video. Our experiments on Hollywood2 show state-of-the-art performance on saliency prediction for video. We also show that our attentional model trained on Hollywood2 generalizes well to UCF101 and it can be leveraged to improve action classification accuracy on both datasets.
Conditional random fields (CRFs) provide a powerful tool for structured prediction, but cast significant challenges in both the learning and inference steps. Approximation techniques are widely used in both steps, which should be considered jointly to guarantee good performance (a.k.a. "inferning"). Perturb-and-MAP models provide a promising alternative to CRFs, but require global combinatorial optimization and hence they are usable only on specific models. In this work, we present a new Local Perturb-and-MAP (locPMAP) framework that replaces the global optimization with a local optimization by exploiting our observed connection between locPMAP and the pseudolikelihood of the original CRF model. We test our approach on three different vision tasks and show that our method achieves consistently improved performance over other approximate inference techniques optimized to a pseudolikelihood objective. Additionally, we demonstrate that we can integrate our method in the fully convolutional network framework to increase our model's complexity. Finally, our observed connection between locPMAP and the pseudolikelihood leads to a novel perspective for understanding and using pseudolikelihood.
The state-of-the-art in semantic segmentation is currently represented by fully convolutional networks (FCNs). However, FCNs use large receptive fields and many pooling layers, both of which cause blurring and low spatial resolution in the deep layers. As a result FCNs tend to produce segmentations that are poorly localized around object boundaries. Prior work has attempted to address this issue in post-processing steps, for example using a color-based CRF on top of the FCN predictions. However, these approaches require additional parameters and low-level features that are difficult to tune and integrate into the original network architecture. Additionally, most CRFs use color-based pixel affinities, which are not well suited for semantic segmentation and lead to spatially disjoint predictions. To overcome these problems, we introduce a Boundary Neural Field (BNF), which is a global energy model integrating FCN predictions with boundary cues. The boundary information is used to enhance semantic segment coherence and to improve object localization. Specifically, we first show that the convolutional filters of semantic FCNs provide good features for boundary detection. We then employ the predicted boundaries to define pairwise potentials in our energy. Finally, we show that our energy decomposes semantic segmentation into multiple binary problems, which can be relaxed for efficient global optimization. We report extensive experiments demonstrating that minimization of our global boundary-based energy yields results superior to prior globalization methods, both quantitatively as well as qualitatively.