Search

Imagery collected from outdoor visual environments is often degraded due to the presence of dense smoke or haze. A key challenge for research in scene understanding in these degraded visual environments (DVE) is the lack of representative benchmark datasets. These datasets are required to evaluate state-of-the-art object recognition and other computer vision algorithms in degraded settings. In this paper, we address some of these limitations by introducing the first realistic haze image benchmark, from both aerial and ground view, with paired haze-free images, and in-situ haze density measurements. This dataset was produced in a controlled environment with professional smoke generating machines that covered the entire scene, and consists of images captured from the perspective of both an unmanned aerial vehicle (UAV) and an unmanned ground vehicle (UGV). We also evaluate a set of representative state-of-the-art dehazing approaches as well as object detectors on the dataset. The full dataset presented in this paper, including the ground truth object classification bounding boxes and haze density measurements, is provided for the community to evaluate their algorithms at: https://a2i2-archangel.vision. A subset of this dataset has been used for the "Object Detection in Haze" Track of CVPR UG2 2022 challenge at https://cvpr2022.ug2challenge.org/track1.html.

We investigate privacy-preserving, video-based action recognition in deep learning, a problem with growing importance in smart camera applications. A novel adversarial training framework is formulated to learn an anonymization transform for input videos such that the trade-off between target utility task performance and the associated privacy budgets is explicitly optimized on the anonymized videos. Notably, the privacy budget, often defined and measured in task-driven contexts, cannot be reliably indicated using any single model performance because strong protection of privacy should sustain against any malicious model that tries to steal private information. To tackle this problem, we propose two new optimization strategies of model restarting and model ensemble to achieve stronger universal privacy protection against any attacker models. Extensive experiments have been carried out and analyzed. On the other hand, given few public datasets available with both utility and privacy labels, the data-driven (supervised) learning cannot exert its full power on this task. We first discuss an innovative heuristic of cross-dataset training and evaluation, enabling the use of multiple single-task datasets (one with target task labels and the other with privacy labels) in our problem. To further address this dataset challenge, we have constructed a new dataset, termed PA-HMDB51, with both target task labels (action) and selected privacy attributes (skin color, face, gender, nudity, and relationship) annotated on a per-frame basis. This first-of-its-kind video dataset and evaluation protocol can greatly facilitate visual privacy research and open up other opportunities. Our codes, models, and the PA-HMDB51 dataset are available at: https://github.com/VITA-Group/PA-HMDB51.

Dedicated equipment, which is widely used in many different types of vehicles, is the core system that determines the combat capability of special vehicles. Therefore, assuring the normal operation of dedicated equipment is crucial. With the increase in battlefield complexity, the demand for equipment functions is increasing, and the complexity of dedicated equipment is also increasing. To solve the problem of fault diagnosis of dedicated equipment, a fault diagnosis algorithm based on CNN-LSTM was proposed in this paper. CNN and LSTM are used in the model adopted by the algorithm to extract spatial and temporal features from the data. CBAM is used to enhance the model's accuracy in identifying faults for dedicated equipment. Data on dedicated equipment faults were obtained from a hardware-in-loop simulation platform to verify the model. It is demonstrated that the proposed fault diagnosis algorithm has high recognition ability for dedicated equipment by comparing it to other neural network models. [ABSTRACT FROM AUTHOR]

• Adaptively allocate computing resource to input regions for better network inference. • Patch sampling algorithm reduces redundant calculation costs in overlapping regions. • Comparable performance is achieved on two datasets by comparing with SOTA methods. [Display omitted] In this paper, we propose an entropy-dynamic resolution detection (EDRdet) method for object detection in aerial images. Most conventional object detection methods usually detect each region in aerial images directly with a fixed resolution, so that the resolution of the hard-to-detect region in the whole aerial image is not enough and that of the easy-to-detect region is not necessary. We argue that different resolutions of regions are required for efficient and accurate object detection in aerial images. Our EDRdet dynamically adjusts the resolution of each region via measuring the challenge of the detection process, therefore, our model can achieve a good trade-off between the computational cost and the accuracy effectively. Specifically, our EDRdet searches hard-to-detect regions through low-resolution global context information, and inputs higher-resolution patches to supplement lost feature at low resolution for fine-grained detection. We apply the entropy to determine the regions in aerial images required to be detected at higher resolutions, and where the entropy represents the uncertainty of the detection result. Moreover, we design a patch sampling algorithm to make the selected regions sparse to further improve the efficiency of patch generation. Extensive experiments on the DOTA and Visdrone2019 datasets verify that our EDRdet can reduce the computational cost and improve the model accuracy effectively. [ABSTRACT FROM AUTHOR]

Data-driven fault diagnostics of industrial systems suffer from class-imbalanced problems, which is a common challenge for machine learning algorithms as it is difficult to learn the features of the minority class samples. Synthetic oversampling methods are commonly used to tackle these problems by generating minority class samples to balance the majority and minority classes. Two major issues will influence the performance of oversampling methods which are how to choose the most appropriate existing minority seed samples, and how to synthesize new samples from seed samples effectively. However, many existing oversampling methods are not accurate and effective enough to generate new samples when dealing with high-dimensional faulty samples with different imbalanced ratios, since they do not take these two factors into consideration at the same time. This article develops a novel adaptive oversampling technique: expectation maximization (EM)-based local-weighted minority oversampling technique for industrial fault diagnostics. This method uses a local-weighted minority oversampling strategy to identify hard-to-learn informative minority fault samples and an EM-based imputation algorithm to generate fault samples based on the distribution of minority samples. To validate the performance of the developed method, experiments were conducted on two real-world datasets. The results show that the developed method can achieve better performances, in terms of F-measure, Matthews correlation coefficient (MCC), and Mean (average of F-measure and MCC) values, on multiclass imbalanced fault diagnostics in different imbalance ratios than state-of-arts’ baseline sampling techniques. [ABSTRACT FROM AUTHOR]