Deep Convolutional Neural Networks (DCNNs) have recently shown state of the art performance in high level vision tasks, such as image classification and object detection. This work brings together methods from DCNNs and probabilistic graphical models for addressing the task of pixel-level classification (also called "semantic image segmentation"). We show that responses at the final layer of DCNNs are not sufficiently localized for accurate object segmentation. This is due to the very invariance properties that make DCNNs good for high level tasks. We overcome this poor localization property of deep networks by combining the responses at the final DCNN layer with a fully connected Conditional Random Field (CRF). Qualitatively, our "DeepLab" system is able to localize segment boundaries at a level of accuracy which is beyond previous methods. Quantitatively, our method sets the new state-of-art at the PASCAL VOC-2012 semantic image segmentation task, reaching 71.6% IOU accuracy in the test set. We show how these results can be obtained efficiently: Careful network re-purposing and a novel application of the 'hole' algorithm from the wavelet community allow dense computation of neural net responses at 8 frames per second on a modern GPU.
Feature representation plays a crucial role in visual correspondence, and recent methods for image matching resort to deeply stacked convolutional layers. These models, however, are both monolithic and static in the sense that they typically use a specific level of features, e.g., the output of the last layer, and adhere to it regardless of the images to match. In this work, we introduce a novel approach to visual correspondence that dynamically composes effective features by leveraging relevant layers conditioned on the images to match. Inspired by both multi-layer feature composition in object detection and adaptive inference architectures in classification, the proposed method, dubbed Dynamic Hyperpixel Flow, learns to compose hypercolumn features on the fly by selecting a small number of relevant layers from a deep convolutional neural network. We demonstrate the effectiveness on the task of semantic correspondence, i.e., establishing correspondences between images depicting different instances of the same object or scene category. Experiments on standard benchmarks show that the proposed method greatly improves matching performance over the state of the art in an adaptive and efficient manner.
The advances and successes in deep learning in recent years have led to considerable efforts and investments into its widespread ubiquitous adoption for a wide variety of applications, ranging from personal assistants and intelligent navigation to search and product recommendation in e-commerce. With this tremendous rise in deep learning adoption comes questions about the trustworthiness of the deep neural networks that power these applications. Motivated to answer such questions, there has been a very recent interest in trust quantification. In this work, we introduce the concept of trust matrix, a novel trust quantification strategy that leverages the recently introduced question-answer trust metric by Wong et al. to provide deeper, more detailed insights into where trust breaks down for a given deep neural network given a set of questions. More specifically, a trust matrix defines the expected question-answer trust for a given actor-oracle answer scenario, allowing one to quickly spot areas of low trust that needs to be addressed to improve the trustworthiness of a deep neural network. The proposed trust matrix is simple to calculate, humanly interpretable, and to the best of the authors' knowledge is the first to study trust at the actor-oracle answer level. We further extend the concept of trust densities with the notion of conditional trust densities. We experimentally leverage trust matrices to study several well-known deep neural network architectures for image recognition, and further study the trust density and conditional trust densities for an interesting actor-oracle answer scenario. The results illustrate that trust matrices, along with conditional trust densities, can be useful tools in addition to the existing suite of trust quantification metrics for guiding practitioners and regulators in creating and certifying deep learning solutions for trusted operation.
Network quantization aims to lower the bitwidth of weights and activations and hence reduce the model size and accelerate the inference of deep networks. Even though existing quantization methods have achieved promising performance on image classification, applying aggressively low bitwidth quantization on object detection while preserving the performance is still a challenge. In this paper, we demonstrate that the poor performance of the quantized network on object detection comes from the inaccurate batch statistics of batch normalization. To solve this, we propose an accurate quantized object detection (AQD) method. Specifically, we propose to employ multi-level batch normalization (multi-level BN) to estimate the batch statistics of each detection head separately. We further propose a learned interval quantization method to improve how the quantizer itself is configured. To evaluate the performance of the proposed methods, we apply AQD to two one-stage detectors (i.e., RetinaNet and FCOS). Experimental results on COCO show that our methods achieve near-lossless performance compared with the full-precision model by using extremely low bitwidth regimes such as 3-bit. In particular, we even outperform the full-precision counterpart by a large margin with a 4-bit detector, which is of great practical value.
Not all errors are created equal. This is especially true for many key machine learning applications. In the case of classification tasks, the hierarchy of errors can be summarized under the form of a cost matrix, which assesses the gravity of confusing each pair of classes. When certain conditions are met, this matrix defines a metric, which we use in a new and versatile classification layer to model the disparity of errors. Our method relies on conjointly learning a feature-extracting network and a set of class representations, or prototypes, which incorporate the error metric into their relative arrangement. Our approach allows for consistent improvement of the network's prediction with regard to the cost matrix. Furthermore, when the induced metric contains insight on the data structure, our approach improves the overall precision. Experiments on three different tasks and public datasets -- from agricultural time series classification to depth image semantic segmentation -- validate our approach.
In this paper, we suggest a new neural network architecture for vanishing point detection in images. The key element is the use of the direct and transposed Fast Hough Transforms separated by convolutional layer blocks with activation functions. It allows us to get the answer in the coordinates of the input image at the output of the network and thus to calculate the coordinates of the vanishing point by simply selecting the maximum. The use of integral operators enables the neural network to rely on global rectilinear features in the image, and so it is ideal for detecting vanishing points. To demonstrate the effectiveness of the proposed architecture, we use a set of images from a DVR and show its superiority over existing methods. Note, in addition, that the proposed neural network architecture essentially repeats the process of direct and back projection used, for example, in computed tomography.
We propose an image representation scheme combining the local and nonlocal characterization of patches in an image. Our representation scheme can be shown to be equivalent to a tight frame constructed from convolving local bases (e.g. wavelet frames, discrete cosine transforms, etc.) with nonlocal bases (e.g. spectral basis induced by nonlinear dimension reduction on patches), and we call the resulting frame elements {\it convolution framelets}. Insight gained from analyzing the proposed representation leads to a novel interpretation of a recent high-performance patch-based image inpainting algorithm using Point Integral Method (PIM) and Low Dimension Manifold Model (LDMM) [Osher, Shi and Zhu, 2016]. In particular, we show that LDMM is a weighted $\ell_2$-regularization on the coefficients obtained by decomposing images into linear combinations of convolution framelets; based on this understanding, we extend the original LDMM to a reweighted version that yields further improved inpainting results. In addition, we establish the energy concentration property of convolution framelet coefficients for the setting where the local basis is constructed from a given nonlocal basis via a linear reconstruction framework; a generalization of this framework to unions of local embeddings can provide a natural setting for interpreting BM3D, one of the state-of-the-art image denoising algorithms.
Deep neural networks demonstrated their ability to provide remarkable performances on a wide range of supervised learning tasks (e.g., image classification) when trained on extensive collections of labeled data (e.g., ImageNet). However, creating such large datasets requires a considerable amount of resources, time, and effort. Such resources may not be available in many practical cases, limiting the adoption and the application of many deep learning methods. In a search for more data-efficient deep learning methods to overcome the need for large annotated datasets, there is a rising research interest in semi-supervised learning and its applications to deep neural networks to reduce the amount of labeled data required, by either developing novel methods or adopting existing semi-supervised learning frameworks for a deep learning setting. In this paper, we provide a comprehensive overview of deep semi-supervised learning, starting with an introduction to the field, followed by a summarization of the dominant semi-supervised approaches in deep learning.
Automated tongue image segmentation in tongue images is a challenging task for two reasons: 1) there are many pathological details on the tongue surface, which affect the extraction of the boundary; 2) the shapes of the tongues captured from various persons (with different diseases) are quite different. To deal with the challenge, a novel end-to-end Boundary Guidance Hierarchical Network (BGHNet) with a new hybrid loss is proposed in this paper. In the new approach, firstly Context Feature Encoder Module (CFEM) is built upon the bottomup pathway to confront with the shrinkage of the receptive field. Secondly, a novel hierarchical recurrent feature fusion module (HRFFM) is adopt to progressively and hierarchically refine object maps to recover image details by integrating local context information. Finally, the proposed hybrid loss in a four hierarchy-pixel, patch, map and boundary guides the network to effectively segment the tongue regions and accurate tongue boundaries. BGHNet is applied to a set of tongue images. The experimental results suggest that the proposed approach can achieve the latest tongue segmentation performance. And in the meantime, the lightweight network contains only 15.45M parameters and performs only 11.22GFLOPS.
Generative models are undoubtedly a hot topic in Artificial Intelligence, among which the most common type is Generative Adversarial Networks (GANs). These architectures let one synthesise artificial datasets by implicitly modelling the underlying probability distribution of a real-world training dataset. With the introduction of Conditional GANs and their variants, these methods were extended to generating samples conditioned on ancillary information available for each sample within the dataset. From a practical standpoint, however, one might desire to generate data conditioned on partial information. That is, only a subset of the ancillary conditioning variables might be of interest when synthesising data. In this work, we argue that standard Conditional GANs are not suitable for such a task and propose a new Adversarial Network architecture and training strategy to deal with the ensuing problems. Experiments illustrating the value of the proposed approach in digit and face image synthesis under partial conditioning information are presented, showing that the proposed method can effectively outperform the standard approach under these circumstances.