Along with the great success of deep neural networks, there is also growing concern about their black-box nature. The interpretability issue affects people's trust on deep learning systems. It is also related to many ethical problems, e.g., algorithmic discrimination. Moreover, interpretability is a desired property for deep networks to become powerful tools in other research fields, e.g., drug discovery and genomics. In this survey, we conduct a comprehensive review of the neural network interpretability research. We first clarify the definition of interpretability as it has been used in many different contexts. Then we elaborate on the importance of interpretability and propose a novel taxonomy organized along three dimensions: type of engagement (passive vs. active interpretation approaches), the type of explanation, and the focus (from local to global interpretability). This taxonomy provides a meaningful 3D view of distribution of papers from the relevant literature as two of the dimensions are not simply categorical but allow ordinal subcategories. Finally, we summarize the existing interpretability evaluation methods and suggest possible research directions inspired by our new taxonomy.

This paper proposes a novel approach to constructing a hierarchical representation of visual input that aims to enable recognition and detection of a large number of object categories. Inspired by the principles of efficient indexing (bottom-up,), robust matching (top-down,), and ideas of compositionality, our approach learns a hierarchy of spatially flexible compositions, i.e. parts, in an unsupervised, statistics-driven manner. Starting with simple, frequent features, we learn the statistically most significant compositions (parts composed of parts), which consequently define the next layer. Parts are learned sequentially, layer after layer, optimally adjusting to the visual data. Lower layers are learned in a category-independent way to obtain complex, yet sharable visual building blocks, which is a crucial step towards a scalable representation. Higher layers of the hierarchy, on the other hand, are constructed by using specific categories, achieving a category representation with a small number of highly generalizable parts that gained their structural flexibility through composition within the hierarchy. Built in this way, new categories can be efficiently and continuously added to the system by adding a small number of parts only in the higher layers. The approach is demonstrated on a large collection of images and a variety of object categories. Detection results confirm the effectiveness and robustness of the learned parts.

Image demoireing is a multi-faceted image restoration task involving both moire pattern removal and color restoration. In this paper, we raise a general degradation model to describe an image contaminated by moire patterns, and propose a novel multi-scale bandpass convolutional neural network (MBCNN) for single image demoireing. For moire pattern removal, we propose a multi-block-size learnable bandpass filters (M-LBFs), based on a block-wise frequency domain transform, to learn the frequency domain priors of moire patterns. We also introduce a new loss function named Dilated Advanced Sobel loss (D-ASL) to better sense the frequency information. For color restoration, we propose a two-step tone mapping strategy, which first applies a global tone mapping to correct for a global color shift, and then performs local fine tuning of the color per pixel. To determine the most appropriate frequency domain transform, we investigate several transforms including DCT, DFT, DWT, learnable non-linear transform and learnable orthogonal transform. We finally adopt the DCT. Our basic model won the AIM2019 demoireing challenge. Experimental results on three public datasets show that our method outperforms state-of-the-art methods by a large margin.

This paper presents a novel robust method for single target tracking in RGB-D images, and also contributes a substantial new benchmark dataset for evaluating RGB-D trackers. While a target object's color distribution is reasonably motion-invariant, this is not true for the target's depth distribution, which continually varies as the target moves relative to the camera. It is therefore nontrivial to design target models which can fully exploit (potentially very rich) depth information for target tracking. For this reason, much of the previous RGB-D literature relies on color information for tracking, while exploiting depth information only for occlusion reasoning. In contrast, we propose an adaptive range-invariant target depth model, and show how both depth and color information can be fully and adaptively fused during the search for the target in each new RGB-D image. We introduce a new, hierarchical, two-layered target model (comprising local and global models) which uses spatio-temporal consistency constraints to achieve stable and robust on-the-fly target relearning. In the global layer, multiple features, derived from both color and depth data, are adaptively fused to find a candidate target region. In ambiguous frames, where one or more features disagree, this global candidate region is further decomposed into smaller local candidate regions for matching to local-layer models of small target parts. We also note that conventional use of depth data, for occlusion reasoning, can easily trigger false occlusion detections when the target moves rapidly toward the camera. To overcome this problem, we show how combining target information with contextual information enables the target's depth constraint to be relaxed. Our adaptively relaxed depth constraints can robustly accommodate large and rapid target motion in the depth direction, while still enabling the use of depth data for highly accurate reasoning about occlusions. For evaluation, we introduce a new RGB-D benchmark dataset with per-frame annotated attributes and extensive bias analysis. Our tracker is evaluated using two different state-of-the-art methodologies, VOT and object tracking benchmark, and in both cases it significantly outperforms four other state-of-the-art RGB-D trackers from the literature.

Data-driven deep learning approaches to image registration can be less accurate than conventional iterative approaches, especially when training data is limited. To address this issue and meanwhile retain the fast inference speed of deep learning, we propose VR-Net, a novel cascaded variational network for unsupervised deformable image registration. Using a variable splitting optimization scheme, we first convert the image registration problem, established in a generic variational framework, into two sub-problems, one with a point-wise, closed-form solution and the other one being a denoising problem. We then propose two neural layers (i.e. warping layer and intensity consistency layer) to model the analytical solution and a residual U-Net (termed generalized denoising layer) to formulate the denoising problem. Finally, we cascade the three neural layers multiple times to form our VR-Net. Extensive experiments on three (two 2D and one 3D) cardiac magnetic resonance imaging datasets show that VR-Net outperforms state-of-the-art deep learning methods on registration accuracy, whilst maintaining the fast inference speed of deep learning and the data-efficiency of variational models.