We propose a new method for creating computationally efficient and compact convolutional neural networks (CNNs) using a novel sparse connection structure that resembles a tree root. This allows a significant reduction in computational cost and number of parameters compared to state-of-the-art deep CNNs, without compromising accuracy, by exploiting the sparsity of inter-layer filter dependencies. We validate our approach by using it to train more efficient variants of state-of-the-art CNN architectures, evaluated on the CIFAR10 and ILSVRC datasets. Our results show similar or higher accuracy than the baseline architectures with much less computation, as measured by CPU and GPU timings. For example, for ResNet 50, our model has 40% fewer parameters, 45% fewer floating point operations, and is 31% (12%) faster on a CPU (GPU). For the deeper ResNet 200 our model has 48% fewer parameters and 27% fewer floating point operations, while maintaining state-of-the-art accuracy. For GoogLeNet, our model has 7% fewer parameters and is 21% (16%) faster on a CPU (GPU).

We propose a new method for creating computationally efficient and compact convolutional neural networks (CNNs) using a novel sparse connection structure that resembles a tree root. This allows a significant reduction in computational cost and number of parameters compared to state-of-the-art deep CNNs, without compromising accuracy, by exploiting the sparsity of inter-layer filter dependencies. We validate our approach by using it to train more efficient variants of state-of-the-art CNN architectures, evaluated on the CIFAR10 and ILSVRC datasets. Our results show similar or higher accuracy than the baseline architectures with much less computation, as measured by CPU and GPU timings. For example, for ResNet 50, our model has 40% fewer parameters, 45% fewer floating point operations, and is 31% (12%) faster on a CPU (GPU). For the deeper ResNet 200 our model has 25% fewer floating point operations and 44% fewer parameters, while maintaining state-of-the-art accuracy. For GoogLeNet, our model has 7% fewer parameters and is 21% (16%) faster on a CPU (GPU).

This paper investigates the connections between two state of the art classifiers: decision forests (DFs, including decision jungles) and convolutional neural networks (CNNs). Decision forests are computationally efficient thanks to their conditional computation property (computation is confined to only a small region of the tree, the nodes along a single branch). CNNs achieve state of the art accuracy, thanks to their representation learning capabilities. We present a systematic analysis of how to fuse conditional computation with representation learning and achieve a continuum of hybrid models with different ratios of accuracy vs. efficiency. We call this new family of hybrid models conditional networks. Conditional networks can be thought of as: i) decision trees augmented with data transformation operators, or ii) CNNs, with block-diagonal sparse weight matrices, and explicit data routing functions. Experimental validation is performed on the common task of image classification on both the CIFAR and Imagenet datasets. Compared to state of the art CNNs, our hybrid models yield the same accuracy with a fraction of the compute cost and much smaller number of parameters.

We propose a new method for creating computationally efficient convolutional neural networks (CNNs) by using low-rank representations of convolutional filters. Rather than approximating filters in previously-trained networks with more efficient versions, we learn a set of small basis filters from scratch; during training, the network learns to combine these basis filters into more complex filters that are discriminative for image classification. To train such networks, a novel weight initialization scheme is used. This allows effective initialization of connection weights in convolutional layers composed of groups of differently-shaped filters. We validate our approach by applying it to several existing CNN architectures and training these networks from scratch using the CIFAR, ILSVRC and MIT Places datasets. Our results show similar or higher accuracy than conventional CNNs with much less compute. Applying our method to an improved version of VGG-11 network using global max-pooling, we achieve comparable validation accuracy using 41% less compute and only 24% of the original VGG-11 model parameters; another variant of our method gives a 1 percentage point increase in accuracy over our improved VGG-11 model, giving a top-5 center-crop validation accuracy of 89.7% while reducing computation by 16% relative to the original VGG-11 model. Applying our method to the GoogLeNet architecture for ILSVRC, we achieved comparable accuracy with 26% less compute and 41% fewer model parameters. Applying our method to a near state-of-the-art network for CIFAR, we achieved comparable accuracy with 46% less compute and 55% fewer parameters.

This paper investigates the connections between two state of the art classifiers: decision forests (DFs, including decision jungles) and convolutional neural networks (CNNs). Decision forests are computationally efficient thanks to their conditional computation property (computation is confined to only a small region of the tree, the nodes along a single branch). CNNs achieve state of the art accuracy, thanks to their representation learning capabilities. We present a systematic analysis of how to fuse conditional computation with representation learning and achieve a continuum of hybrid models with different ratios of accuracy vs. efficiency. We call this new family of hybrid models conditional networks. Conditional networks can be thought of as: i) decision trees augmented with data transformation operators, or ii) CNNs, with block-diagonal sparse weight matrices, and explicit data routing functions. Experimental validation is performed on the common task of image classification on both the CIFAR and Imagenet datasets. Compared to state of the art CNNs, our hybrid models yield the same accuracy with a fraction of the compute cost and much smaller number of parameters.

This paper investigates the connections between two state of the art classifiers: decision forests (DFs, including decision jungles) and convolutional neural networks (CNNs). Decision forests are computationally efficient thanks to their conditional computation property (computation is confined to only a small region of the tree, the nodes along a single branch). CNNs achieve state of the art accuracy, thanks to their representation learning capabilities. We present a systematic analysis of how to fuse conditional computation with representation learning and achieve a continuum of hybrid models with different ratios of accuracy vs. efficiency. We call this new family of hybrid models conditional networks. Conditional networks can be thought of as: i) decision trees augmented with data transformation operators, or ii) CNNs, with block-diagonal sparse weight matrices, and explicit data routing functions. Experimental validation is performed on the common task of image classification on both the CIFAR and Imagenet datasets. Compared to state of the art CNNs, our hybrid models yield the same accuracy with a fraction of the compute cost and much smaller number of parameters.

We propose a new method for creating computationally efficient convolutional neural networks (CNNs) by using low-rank representations of convolutional filters. Rather than approximating filters in previously-trained networks with more efficient versions, we learn a set of small basis filters from scratch; during training, the network learns to combine these basis filters into more complex filters that are discriminative for image classification. To train such networks, a novel weight initialization scheme is used. This allows effective initialization of connection weights in convolutional layers composed of groups of differently-shaped filters. We validate our approach by applying it to several existing CNN architectures and training these networks from scratch using the CIFAR, ILSVRC and MIT Places datasets. Our results show similar or higher accuracy than conventional CNNs with much less compute. Applying our method to an improved version of VGG-11 network using global max-pooling, we achieve comparable validation accuracy using 41% less compute and only 24% of the original VGG-11 model parameters; another variant of our method gives a 1 percentage point increase in accuracy over our improved VGG-11 model, giving a top-5 center-crop validation accuracy of 89.7% while reducing computation by 16% relative to the original VGG-11 model. Applying our method to the GoogLeNet architecture for ILSVRC, we achieved comparable accuracy with 26% less compute and 41% fewer model parameters. Applying our method to a near state-of-the-art network for CIFAR, we achieved comparable accuracy with 46% less compute and 55% fewer parameters.

Deep Convolutional Neural Networks (CNNs) have recently evinced immense success for various image recognition tasks [11, 27]. However, a question of paramount importance is somewhat unanswered in deep learning research - is the selected CNN optimal for the dataset in terms of accuracy and model size?In this paper, we intend to answer this question and introduce a novel strategy that alters the architecture of a given CNN for a specified dataset, to potentially enhance the original accuracy while possibly reducing the model size. We use two operations for architecture refinement, viz. stretching and symmetrical splitting. Stretching increases the number of hidden units (nodes) in a given CNN layer, while a symmetrical split of say K between two layers separates the input and output channels into K equal groups, and connects only the corresponding input-output channel groups. Our procedure starts with a pre-trained CNN for a given dataset, and optimally decides the stretch and split factors across the network to refine the architecture. We empirically demonstrate the necessity of the two operations.We evaluate our approach on two natural scenes attributes datasets, SUN Attributes [16] and CAMIT-NSAD [20], with architectures of GoogleNet and VGG-11, that are quite contrasting in their construction. We justify our choice of datasets, and show that they are interestingly distinct from each other, and together pose a challenge to our architectural refinement algorithm. Our results substantiate the usefulness of the proposed method.

We present a new real-time hand tracking system based on a single depth camera. The system can accurately reconstruct complex hand poses across a variety of subjects. It also allows for robust tracking, rapidly recovering from any temporary failures. Most uniquely, our tracker is highly flexible, dramatically improving upon previous approaches which have focused on front-facing close-range scenarios. This flexibility opens up new possibilities for human-computer interaction with examples including tracking at distances from tens of centimeters through to several meters (for controlling the TV at a distance), supporting tracking using a moving depth camera (for mobile scenarios), and arbitrary camera placements (for VR headsets). These features are achieved through a new pipeline that combines a multi-layered discriminative reinitialization strategy for per-frame pose estimation, followed by a generative model-fitting stage. We provide extensive technical details and a detailed qualitative and quantitative analysis.

This chapter discusses the use of regression forests for the automatic detection and simultaneous localization of multiple anatomical regions within computed tomography (CT) and magnetic resonance (MR) three-dimensional images. Important applications include: organ-specific tracking of radiation dose over time; selective retrieval of patient images from radiological database systems; semantic visual navigation; and the initialization of organ-specific image processing operations. We present a continuous parametrization of the anatomy localization problem, which allows it to be addressed effectively by multivariate random regression forests (Chap. 5 ). A single pass of our probabilistic algorithm enables the direct mapping from voxels to organ location and size, with training focusing on maximizing the confidence of output predictions. As a by-product, our method produces salient anatomical landmarks, i.e. automatically selected "anchor" regions which help localize organs of interest with high confidence. This chapter builds upon the work in Criminisi et al., in MICCAI workshop on medical computer vision: recognition techniques and applications in medical imaging, 2010 and in Pauly et al., Proc medical image computing and computer assisted intervention, 2011 and demonstrates the flexibility of forests in dealing with both CT and multi-channel MR scans. Quantitative validation is performed on two ground truth labeled datasets: (i) a database of 400 highly variable CT scans, and (ii) a database of 33 full-body, multi-channel MR scans. In both cases localization errors are reduced and results are more stable than those from more conventional atlas-based registration approaches. The simplicity of the regressor's context-rich visual features yield typical run-times of only 4 seconds per scan on a standard desktop. This anatomy recognition algorithm has now received FDA approval and is part of Caradigm's Amalga ( www.caradigm.com ).

This paper proposes a new algorithm for the efficient, automatic detection and localization of multiple anatomical structures within three-dimensional computed tomography (CT) scans. Applications include selective retrieval of patients images from PACS systems, semantic visual navigation and tracking radiation dose over time. The main contribution of this work is a new, continuous parametrization of the anatomy localization problem, which allows it to be addressed effectively by multi-class random regression forests. Regression forests are similar to the more popular classification forests, but trained to predict continuous, multi-variate outputs, where the training focuses on maximizing the confidence of output predictions. A single pass of our probabilistic algorithm enables the direct mapping from voxels to organ location and size. Quantitative validation is performed on a database of 400 highly variable CT scans. We show that the proposed method is more accurate and robust than techniques based on efficient multi-atlas registration and template-based nearest-neighbor detection. Due to the simplicity of the regressor's context-rich visual features and the algorithm's parallelism, these results are achieved in typical run-times of only ∼4 s on a conventional single-core machine.

Existing literature describes a variety of techniques for semantic annotation of DICOM CT images, i.e. the automatic detection and localization of anatomical structures. Semantic annotation facilitates enhanced image navigation, linkage of DICOM image content and non-image clinical data, content-based image retrieval, and image registration. A key challenge for semantic annotation algorithms is inter-patient variability. However, while the algorithms described in published literature have been shown to cope adequately with the variability in test sets comprising adult CT scans, the problem presented by the even greater variability in pediatric anatomy has received very little attention. Most existing semantic annotation algorithms can only be extended to work on scans of both adult and pediatric patients by adapting parameters heuristically in light of patient size. In contrast, our approach, which uses random regression forests ('RRF'), learns an implicit model of scale variation automatically using training data. In consequence, anatomical structures can be localized accurately in both adult and pediatric CT studies without the need for parameter adaptation or additional information about patient scale. We show how the RRF algorithm is able to learn scale invariance from a combined training set containing a mixture of pediatric and adult scans. Resulting localization accuracy for both adult and pediatric data remains comparable with that obtained using RRFs trained and tested using only adult data.

This paper describes an interactive system for quickly modelling 3D body shapes from a single image. It provides the user with a convenient way to obtain their 3D body shapes so as to try on virtual garments online. For the ease of use, we first introduce a novel interface for users to conveniently extract anthropometric measurements from a single photo, while using readily available scene cues for automatic image rectification. Then, we propose a unified probabilistic framework using Gaussian processes, which predict the body parameters from input measurements while correcting the aspect ratio ambiguity resulting from photo rectification. Extensive experiments and user studies have supported the efficacy of our system. This system is now being exploited commercially online(1).

In the current health-care environment, the time available for physicians to browse patients' scans is shrinking due to the rapid increase in the sheer number of images. This is further aggravated by mounting pressure to become more productive in the face of decreasing reimbursement. Hence, there is an urgent need to deliver technology which enables faster and effortless navigation through sub-volume image visualizations. Annotating image regions with semantic labels such as those derived from the RADLEX ontology can vastly enhance image navigation and sub-volume visualization. This paper uses random regression forests for efficient, automatic detection and localization of anatomical structures within DICOM 3D CT scans. A regression forest is a collection of decision trees which are trained to achieve direct mapping from voxels to organ location and size in a single pass. This paper focuses on comparing automated labeling with expert-annotated ground-truth results on a database of 50 highly variable CT scans. Initial investigations show that regression forest derived localization errors are smaller and more robust than those achieved by state-of-the-art global registration approaches. The simplicity of the algorithm's context-rich visual features yield typical runtimes of less than 10 seconds for a 5123 voxel DICOM CT series on a single-threaded, single-core machine running multiple trees; each tree taking less than a second. Furthermore, qualitative evaluation demonstrates that using the detected organs' locations as index into the image volume improves the efficiency of the navigational workflow in all the CT studies.

Global linear registration is a necessary first step for many different tasks in medical image analysis. Comparing longitudinal studies1, cross-modality fusion2, and many other applications depend heavily on the success of the automatic registration. The robustness and efficiency of this step is crucial as it affects all subsequent operations. Most common techniques cast the linear registration problem as the minimization of a global energy function based on the image intensities. Although these algorithms have proved useful, their robustness in fully automated scenarios is still an open question. In fact, the optimization step often gets caught in local minima yielding unsatisfactory results. Recent algorithms constrain the space of registration parameters by exploiting implicit or explicit organ segmentations, thus increasing robustness4,5. In this work we propose a novel robust algorithm for automatic global linear image registration. Our method uses random regression forests to estimate posterior probability distributions for the locations of anatomical structures - represented as axis aligned bounding boxes6. These posterior distributions are later integrated in a global linear registration algorithm. The biggest advantage of our algorithm is that it does not require pre-defined segmentations or regions. Yet it yields robust registration results. We compare the robustness of our algorithm with that of the state of the art Elastix toolbox7. Validation is performed via 1464 pair-wise registrations in a database of very diverse 3D CT images. We show that our method decreases the "failure" rate of the global linear registration from 12.5% (Elastix) to only 1.9%.