Recently, stereo matching models based on 3D CNNs have achieved excellent performance. However, huge computational burden and memory footprint of deep 3D convolution limit their deployment on edge devices and real-time scenes. Furthermore, given that disparity maps exhibit continuous disparity variations in content regions but sharp disparity variants at edge areas, spatially shared weight mechanism inherent in convolution may struggle to handle both types of regions. In this paper, we propose a lightweight stereo matching model, called DBCANet, with Dual-Branch Cost Aggregation based on region segmentation to reduce the computational burden. Specifically, we divide the cost aggregation module into two branches: the edge branch is to aggregate edge disparity information, and the content branch is to aggregate the disparity information of the content region. 2D convolution reduces the computational burden and dual branch can mitigate the problem caused by convolutional spatially-shared weights to compensate for the accuracy loss. We extensively validate our model on Scene Flow, KITTI datasets, and our model can achieve the lowest multiply-accumulate operations (MACs) among speed-oriented stereo matching models, while its EPE is only 0.59 on Scene Flow dataset, which is ranked second. Our model can achieve a good balance between accuracy and model computation.

Unsupervised Domain Adaptation for image classification aims to adapt models trained on a labeled source domain to an unlabeled target domain, improving target domain classification performance. However, previous methods often focus solely on sample-level relationships, neglecting the problem of redundancy within samples due to the same feature information being expressed repeatedly. This oversight may reduce the effectiveness of domain adaptation. To address this problem, we propose Redundancy Optimization via Mutual Information for Unsupervised Domain Adaptation (ROMI) to enhance UDA by mitigating feature redundancy. Specifically, we first utilize mutual information to evaluate the redundancy between different feature dimensions and minimize this dimensional mutual information, thereby reducing the redundancy percentage in the features. Subsequently, we maximize the dimensional mutual information between pairs of positive samples in a contrastive learning framework to achieve a more unbiased feature representation. Our extensive experiments on three well-recognized UDA vision benchmarks provide compelling evidence of the efficacy of ROMI.

In adversarial training, it is important to utilize the limited expressive ability of the model to enhance the learning of vulnerable data points. Conventional adversarial training treats all adversarial data equally in the loss function, which may result in unreasonable allocation of the model’s expressive ability, limiting further robustness enhancement. We propose a dynamic weighting loss method based on robust distance to adjust the decision boundary of the model at data points. It enables the model to prioritize the robustness improvement of the more vulnerable data points in adversarial training. The method quantifies the robustness of the model over data points by robust distances, defining relative distances as adjusted expectations of the model’s decision boundaries at the data points and calculating the adversarial learning priorities for the data points. Experimental results on multiple datasets show that the method can further improve the robustness of the model.

Domain Adaptive Object Detection (DAOD) aims to adapt a detector trained on the labeled source domain to perform effectively on the unlabeled target domain. Most existing DAOD methods commonly employ self-training strategy and adversarial learning. However, the methods based on the self-training strategy often generate unreliable pseudo-labels (e.g., missing detections and false positives) due to the absence of target domain semantics, resulting in suboptimal models. Meanwhile, aligning a large number of unimportant pixels hampers adversarial learning’s ability to capture domain-invariant semantic information. Therefore, we propose the Multi-scale Adversarial Teacher Framework (MATF) based on the common self-training framework (teacher–student framework). Specifically, to mitigate the impact of unreliable pseudo-labels, we propose the Multi-scale Negatives Filtering (MNF) module in the teacher model, which prevents missing detections from damaging the model by filtering out unreliable negative samples at multiple feature scales. In addition, we propose the Multi-scale Key-Pixel Discriminator (MKPD), which predicts the distributions of key pixels at multiple feature scales emphasizing the adversarial alignment of key pixels rather than unimportant pixels. Finally, to generate higher-quality pseudo-labels, we introduce the discriminator into the student model to reduce the model’s bias towards the source domain. Unlike most methods, we construct our method on a computationally efficient but less work-intensive one-stage detector. Extensive experiments conducted on DAOD benchmark datasets such as Cityscapes, FoggyCityscapes, KITTI, and Sim10k demonstrate the strong adaptability and effectiveness of MATF.

Unsupervised domain adaptation (UDA) aims to transfer knowledge from the labeled source domain to the fully unlabeled target domain, thus improving the classification performance of the target domain. Recently, self-training methods have shown their effectiveness on UDA. It iteratively trains target data using the generated target pseudo-labels. However, the feature space for generating pseudo-labels contains a large amount of source information, which traps the model in the source domain, making it challenging for the generator to learn discriminative features of the target domain. In this article, we propose a self-training domain adaptation (DA) model with bidirectional transmission generators (BDTGs). Specifically, we design a bidirectional transmission structure for generators, using exponential moving average (EMA) as the bridge between two generators. The structure has two advantages: 1) by transmitting weight parameters to each other during the training process, it promotes the shift of the feature space, thereby alleviating the difficulty of the model in adapting target domain features and 2) the transmission disturbs the classification boundary and is able to expose unreliable target samples near the boundary. We design a cosine similarity-based filter to identify such samples, to reduce the influence of noisy pseudo-labels with incorrect semantic information on the model. Extensive experiments conducted on five benchmark UDA datasets show that our approach has superior classification performance.

Unsupervised domain adaptation (UDA) aims to transfer knowledge from the labeled source domain to the fully-unlabeled target domain, thus improving the classification performance of the target domain. Recently, self-training has shown its effectiveness on UDA. However, the feature space for generating pseudo-labels contains a large amount of source information, making it challenging for the generator to learn discriminative features of the target domain. In this paper, we propose a self-training domain adaptation model via weight transmission between generators (WTBG). Specifically, we develop a bi-directional transmission structure for generators, using Exponential Moving Average (EMA) as the bridge between two generators. By cyclically transmitting weight parameters between them, alleviate the difficulty of generators in learning target features. And a pseudo-label filter based on cosine similarity is designed to reduce the influence of error pseudo-labels. Extensive experiments conducted on two benchmark UDA datasets show that WTBG has superior classification performance.

Unsupervised domain adaptation (UDA) aims to alleviate the problem of distribution differences between unlabeled target domain and labeled source domain. Multi-target domain adaptation (MTDA) requires simultaneous transfer of knowledge from a single source domain to multiple target domains. Although the mixed-domain methods and some solutions that integrate single-target domain adaptation (STDA) optimization strategies have been proposed, the problem of knowledge forgetting has not been well addressed. To solve this problem, this paper proposes a multi-target domain adaptation incremental method based on contrastive learning. We construct a contrastive adaptation network and an incremental container network, which are connected by a feature discriminator. Contrastive learning ensures the adaptation effect of each source-target domain pair. Incremental module saves the weights of contrastive module, and the both networks combine feature discriminator to align the source features, alleviating the knowledge forgetting proplem. Extensive experiments on public datasets demonstrate that our method achieves excellent classification performance.

Network binarization (i.e., binary neural networks, BNNs) can efficiently compress deep neural networks and accelerate model inference but cause severe accuracy degradation. Existing BNNs are mainly implemented based on the commonly used full-precision network backbones, and then the accuracy is improved with various techniques. However, there is a question of whether the full-precision network backbone is well adapted to BNNs. We start from the factors of the performance degradation of BNNs and analyze the problems of directly using full-precision network backbones for BNNs: for a given computational budget, the backbone of a BNN may need to be shallower and wider compared to the backbone of a full-precision network. With this in mind, Depth-Width Reshaping (DWR) is proposed to reshape the depth and width of existing full-precision network backbones and further optimize them by incorporating pruning techniques to better fit the BNNs. Extensive experiments demonstrate the analytical result and the effectiveness of the proposed method. Compared with the original backbones, the DWR backbones constructed by the proposed method result in close to O(√s) decrease in activations, while achieving an absolute accuracy increase by up to 1.7% with comparable computational cost. Besides, by using the DWR backbones, existing methods can achieve new state-of-the-art (SOTA) accuracy (e.g., 67.2% on ImageNet with ResNet-18 as the original backbone). We hope this work provides a novel insight into the backbone design of BNNs. The code is available at https://github.com/pingxue-hfut/DWR.

Unsupervised domain adaptation aims to transfer the knowledge learned from the labeled source domain to the unlabeled target domain, thereby improving the classification performance of the target domain. Recent methods use contrastive learning to optimize this task, however, these methods only focus on contrastive learning for aspects of domain alignment, which do not actively serve the classification task, resulting in suboptimal solution. To this end, we proposed contrastive domain adaptation via delimitation discriminator (CDVD), which addresses the inconsistency problem of optimizing contrastive learning and classification tasks. We introduce a delimitation discriminator to maximize the output difference between two types of random data-augmented samples of the target domain to detect data-augmented samples that are not conducive to classification, and then minimize the difference through the feature generator to generate effective data-augmented features (good for classification). Our extensive experiments on several public datasets show that CDVD has good adaptability and effectiveness.

The unsupervised domain adaptive classification task can learn domain-invariant features between the unlabeled target domain data and the labeled source domain data, thereby improving the classification performance of the classifier in the target domain. However, privacy protection and memory-constrained often make it difficult to obtain labeled source domain samples, which will bring bottlenecks to the existing domain adaptation tasks. To this end, we propose a novel source free domain adaptive classification model, that is, without any source domain data, a classifier with good performance in the target domain can be obtained only by using the source domain pre-trained classifier and the target domain data. The method first proposes a novel conditional information generative adversarial module based on combined discriminators. Through the confrontation between combined discriminators and the generator, the middle domain with pseudo-labels is generated to solve the problem of missing source domain. Then when training the new classifier in the domain adaptation module, we add a distillation loss mechanism to deal with the lack of source domain data supervision, thereby minimizing the difference between the old classifier response and the new classifier response to ensure that the network output retains the source domain information. We conducted experiments on three groups of 10 data sets, which proved that our method can effectively solve the problem of source free domain adaptive classification and effectively improve the classification accuracy of the model in each domain.

The state-of-the-art object detection frameworks often suffer from a performance decline when the feature distribution differences between the source (training) domain and the target (testing) domain are existed. To alleviate this problem, recent works proposed various domain adaptation methods to improve object detection frameworks. Existing methods only consider the feature discrepancy between the source and target domains and ignores unbalanced in class space under cross-domain settings, which can lead to serious negative transfer problems. To address this issue, we propose a novel domain adaptation for object detection to reduce class space discrepancy between domains. Specifically, the weighted mechanism is used to increase the weight of public categories between domains to promote positive transfer and reduce the weight of non-public categories to retard the impact of negative transfer. Moreover, the model reduces domain feature distribution discrepancy by adding domain classifiers and employing adversarial training methods. The results of our experiments on several datasets demonstrate that our model can effectively solve the problem of performance degradation caused by the discrepancy in class space and significantly improve the detection accuracy in each domain.

The field of conversation generation using neural networks has attracted increasing attention from researchers for several years. However, traditional neural language models tend to generate a generic reply with poor semantic logic and no emotion. This article proposes an emotional conversation generation model based on a Bayesian deep neural network that can generate replies with rich emotions, clear themes, and diverse sentences. The topic and emotional keywords of the replies are pregenerated by introducing commonsense knowledge in the model. The reply is divided into multiple clauses, and then a multidimensional generator based on the transformer mechanism proposed in this article is used to iteratively generate clauses from two dimensions: sentence granularity and sentence structure. Subjective and objective experiments prove that compared with existing models, the proposed model effectively improves the semantic logic and emotional accuracy of replies. This model also significantly enhances the diversity of replies, largely overcoming the shortcomings of traditional models that generate safe replies.

Due to the increasing usage of cloud computing applications, it is important to minimize energy cost consumed by a data center, and simultaneously, to improve quality of service via data center management. One promising approach is to switch some servers in a data center to the idle mode for saving energy while to keep a suitable number of servers in the active mode for providing timely service. In this paper, we design both online and offline algorithms for this problem. For the offline algorithm, we formulate data center management as a cost minimization problem by considering energy cost, delay cost (to measure service quality), and switching cost (to change servers’s active/idle mode). Then, we analyze certain properties of an optimal solution which lead to a dynamic programming based algorithm. Moreover, by revising the solution procedure, we successfully eliminate the recursive procedure and achieve an optimal offline algorithm with a polynomial complexity. For the online algorithm, We design it by considering the worst case scenario for future workload. In simulation, we show this online algorithm can always provide near-optimal solutions.