查看更多>>摘要：This paper focuses on high-transferable adversarial attacks on detectors, which are hard to attack in a black-box manner, because of their multiple-output characteristics and the diversity across architectures. To pursue a high attack transferability, one plausible way is to find a common property across detectors, which facilitates the discovery of common weaknesses. We are the first to suggest that the relevance map from interpreters for detectors is such a property. Based on it, we design a Relevance Attack on Detectors (RAD), which achieves a state-of-the-art transferability, exceeding existing results by above 20%. On MS COCO, the detection mAPs for all 8 black-box architectures are more than halved and the segmentation mAPs are also significantly influenced. Given the great transferability of RAD, we generate the first adversarial dataset for object detection and instance segmentation, i.e., Adversarial Objects in COntext (AOCO), which helps to quickly evaluate and improve the robustness of detectors. (c) 2021 Elsevier Ltd. All rights reserved.
查看更多>>摘要：A weakly-supervised learning framework named as complementary-label learning has been proposed re-cently, where each sample is equipped with a single complementary label that denotes one of the classes the sample does not belong to. However, the existing complementary-label learning methods cannot learn from the easily accessible unlabeled samples and samples with multiple complementary labels, which are more informative. In this paper, to remove these limitations, we propose the novel multi-complementary and unlabeled learning framework that allows unbiased estimation of classification risk from samples with any number of complementary labels and unlabeled samples, for arbitrary loss functions and models. We first give an unbiased estimator of the classification risk from samples with multiple complementary labels, and then further improve the estimator by incorporating unlabeled samples into the risk formu-lation. The estimation error bounds show that the proposed methods are in the optimal parametric con-vergence rate. We also propose a risk correction scheme for alleviating over-fitting caused by negative empirical risk. Finally, the experiments on both linear and deep models show the effectiveness of our proposed methods. (c) 2021 Elsevier Ltd. All rights reserved.
查看更多>>摘要：Label noise significantly degrades the generalization ability of deep models in applications. Effective strategies and approaches ( e.g. , re-weighting or loss correction) are designed to alleviate the negative impact of label noise when training a neural network. Those existing works usually rely on the prespecified architecture and manually tuning the additional hyper-parameters. In this paper, we propose warped probabilistic inference (WarPI) to achieve adaptively rectifying the training procedure for the classification network within the meta-learning scenario. In contrast to the deterministic models, WarPI is formulated as a hierarchical probabilistic model by learning an amortization meta-network, which can resolve sample ambiguity and be therefore more robust to serious label noise. Unlike the existing approximated weighting function of directly generating weight values from losses, our meta-network is learned to estimate a rectifying vector from the input of the logits and labels, which has the capability of leveraging sufficient information lying in them. The procedure provides an effective way to rectify the learning procedure for the classification network, demonstrating a significant improvement of the generalization ability. Besides, modeling the rectifying vector as a latent variable and learning the meta-network can be seamlessly integrated into the SGD optimization of the classification network. We evaluate WarPI on four benchmarks of robust learning with noisy labels and achieve the new state-of-the-art under variant noise types. Extensive study and analysis also demonstrate the effectiveness of our model. (c) 2021 Elsevier Ltd. All rights reserved.
查看更多>>摘要：The reflection effect is unavoidable when taking photos through glasses or other transparent materials, which introduces undesired information into pictures. Hence, removing the influence of reflection becomes a key problem in computer vision. One of the main obstacles of recent learning based approaches is the lacking of realistic training data. To address this issue, we introduce a new dataset synthesis method as well as a novel neural network architecture for single image reflection removal. First, we make use of the polarization characteristics of light into the synthesis of datasets, so as to obtain more realistic and diversified training dataset POL. Then, we design a novel Progressive Polarization based Reflection Removal Network ((PR2)-R-2 Net), which preliminary estimates the coarse background layer to guide the final reflection removal. We demonstrate that our method performs better than the state-of-the-art single image reflection removal methods through quantitative and qualitative experimental comparisons. Specifically, the average PSNR of our restored images selected from three representative benchmark datesets: "Real20", "SIR2" and "Nature" is improved at least 0.49 compared with existing methods and reaches to 24.52. (C) 2021 Elsevier Ltd. All rights reserved.
查看更多>>摘要：Image segmentation is a very important topic in the field of computer vision. We present a method for semantic segmentation of selected stuff classes from a superset of classes. We show that in situations where only select stuff classes are required if we group them as per a strategy then it can attain much higher accuracy than the models trained on the original dataset with all classes intact. The COCO-Stuff Dataset is used for demonstrating the aforesaid strategy. For training purposes, the DeepLabv3+ with Mobilenet-v2 architecture is used. We have achieved an 80.2 percent mean Intersection over Union (mIoU) on these selected classes. We also refine the masks using Learning/Computer Vision (CV) methods and hence obtain better visualization results as compared to the existing DeepLabv3+ results. (c) 2021 Elsevier Ltd. All rights reserved.
查看更多>>摘要：Characterizing signal dynamics with network approaches have attracted significant attention in nonlinear time series analysis. Among these approaches, ordinal networks have received great interest for their simplicity and computational efficiency. But most studies mainly use the topological structure of ordinal network to characterize time series while the underlying information in the transition probabilities remain insufficiently concerned. In this paper, the authors introduce an ordinal network-based complexity entropy curve to fill this gap. The numerical results show that this curve has a great discriminating power for signals with different dynamics, outperforming the recently proposed global node entropy. In the empirical application on stock indices, these curves distinguish stock market with different market development and further identify the impact of the 2008 global financial crisis on stock market dynamics. In the analysis of geomagnetic activity, these curves detect the dynamical change in Earths magnetic field caused by the geomagnetic storm. (c) 2021 Elsevier Ltd. All rights reserved.
查看更多>>摘要：The Synthetic Minority Over-sampling Technique (SMOTE) is a well-known resampling strategy that has been successfully used for dealing with the class-imbalance problem, one of the most challenging pattern recognition tasks in the last two decades. In this work, we claim that SMOTE has an important issue when defining the neighborhood in order to create new minority samples: the use of the Euclidean distance may not be suitable in high-dimensional settings. Our hypothesis is that the use of a weighted metric that does not assume that all features are equally important could improve performance in the presence of noisy/redundant variables. In this line, we present a novel SMOTE-like method that uses the weighted Minkowski distance for defining the neighborhood for each example of the minority class. This methodology leads to a better definition of the neighborhood since it prioritizes those features that are more relevant for the classification task. A complementary advantage of the proposal is performing feature selection since attributes can be discarded when their corresponding weights are below a given threshold. Our experiments on 42 class-imbalance datasets show the virtues of the proposed SMOTE variant, achieving the best predictive performance when compared with the traditional SMOTE approach and other recent variants on low-and high-dimensional settings, handling issues such as class overlap and hubness adequately without increasing the complexity of the method. (c) 2021 Elsevier Ltd. All rights reserved.
查看更多>>摘要：Person re-identification (Re-ID) aims to search for the same pedestrian in different cameras, which is a crucial research direction in pattern recognition. Recent deep learning methods have advanced the development of Re-ID. However, the existing approaches easily result in performance degradation in the case of larger scene data because they do not adequately consider the spatial dependencies of both the inter-image and the intra-image. The paper proposes a novel Spatial-Driven Network (SDN) to learn particularly discriminative features with abundant semantic information from both the inter-image and the intra-image dependencies for person Re-ID. Firstly, we design a global-correlation attention module to capture the inter-image dependencies among a series of different pedestrian images. Secondly, we present a local-correlation attention module to compute the intra-image dependencies from any pair of pixels within each pedestrian image. Furthermore, we propose a specific network integration mechanism, which carefully combines the above two complementary modules to match well the solution of the spatial dependency problem. We implement numerous experiments to assess the proposed SDN on mainstream person Re-ID databases. The results demonstrate that the proposed SDN outperforms most of the stateof-the-art methods in typical key criteria. (c) 2021 Elsevier Ltd. All rights reserved.
查看更多>>摘要：Most modern face completion approaches adopt an autoencoder or its variants to restore missing regions in face images. Encoders are often utilized to learn powerful representations that play an important role in meeting the challenges of sophisticated learning tasks. Specifically, various kinds of masks are often presented in face images in the wild, forming complex patterns, especially in this hard period of COVID19. It's difficult for encoders to capture such powerful representations under this complex situation. To address this challenge, we propose a self-supervised Siamese inference network to improve the generalization and robustness of encoders. It can encode contextual semantics from full-resolution images and obtain more discriminative representations. To deal with geometric variations of face images, a dense correspondence field is integrated into the network. We further propose a multi-scale decoder with a novel dual attention fusion module (DAF), which can combine the restored and known regions in an adaptive manner. This multi-scale architecture is beneficial for the decoder to utilize discriminative representations learned from encoders into images. Extensive experiments clearly demonstrate that the proposed approach not only achieves more appealing results compared with state-of-the-art methods but also improves the performance of masked face recognition dramatically. (c) 2021 Elsevier Ltd. All rights reserved.
查看更多>>摘要：Although deep learning-based stereo matching approaches have achieved excellent performance in recent years, it is still a non-trivial task to estimate the uncertainty of the produced disparity map. In this paper, we propose a novel approach to estimate both aleatoric and epistemic uncertainties for stereo matching in an end-to-end way. We introduce an evidential distribution, named Normal Inverse-Gamma (NIG) distribution, whose parameters can be used to calculate the uncertainty. Instead of directly regressed from aggregated features, the uncertainty parameters are predicted for each potential disparity and then averaged via the guidance of matching probability distribution. Furthermore, considering the sparsity of ground truth in real scene datasets, we design two additional losses. The first one tries to enlarge uncertainty on incorrect predictions, so uncertainty becomes more sensitive to erroneous regions. The second one enforces the smoothness of the uncertainty in the regions with smooth disparity. Most stereo matching models, such as PSM-Net, GA-Net, and AA-Net, can be easily integrated with our approach. Experiments on multiple benchmark datasets show that our method improves stereo matching results. We prove that both aleatoric and epistemic uncertainties are well-calibrated with incorrect predictions. Particularly, our method can capture increased epistemic uncertainty on out-of-distribution data, making it effective to prevent a system from potential fatal consequences. Code is available at https://github.com/Dawnstar8411/StereoMatching-Uncertainty . (c) 2021 Elsevier Ltd. All rights reserved.
NSTL
Elsevier