Standard domain adaptation methods require access to both source and target data. However, sharing source data is often impractical in real-world scenarios due to data privacy and memory limitation issues. In this work, we focus on the source-free domain adaptation (SFDA). Existing SFDA methods mainly learn independent information within individual samples and lack the utilization of topological information between samples. For this reason, we explicitly constrain the sample relational knowledge in the mean teacher framework for solving SFDA. Specifically, three relational graphs are first constructed based on the similarity between sample feature pairs: teacher, student, and teacher-student. Then, model adaptation is achieved via two consistency regularizations: 1) Inter-graph consistency constrains the consistency between graph structures. 2) Intra-graph consistency enhances the compactness of the samples within classes. In addition, to mitigate the effect of noisy pseudo labels, local prototypes during the iterations are continuously utilized to calibrate the global prototype to generate high-quality pseudo labels for the target samples. Further, the classification loss is reweighted according to the uncertainty of the pseudo labels, which allows the model to not only highlight the role of high-reliability samples but also to fully exploit the entire target domain. Extensive experimental results on Office-31, Office-Home, VisDA-2017, DomainNet and Digit dataset demonstrate the effectiveness of our method.

Object detection in aerial images is a challenging task for two main reasons: small object and scale variation. Existing methods utilize multi-level features to solve the scale variation problem, but ignore the scale confusion problem of shallow features, limiting the small object detection performance. To solve this issue, we propose a scale decoupling module to emphasizes small object features by eliminating large object features in shallow layers. Moreover, a sparse non-local attention (SNLA) and an adaptive anchor matching strategy (AAMS) are proposed to further improve the small object detection performance. The SNLA only aggregates contextual information of specific sparse positions, which not only refines small object features but also is computationally friendly. The AAMS is suitable for the measurement of small objects, and it can assign more positive samples to small objects. Extensive experiments on 3 challenging aerial datasets, VisDrone-DET2019, UAVDT and DIOR, demonstrate the effectiveness and adaptivity of our method. Code will be available online (https://github.com/MaYou1997/SDP).

Difficult object detection and class imbalance in object detection are the two main challenges faced by aerial image object detection. Difficult objects include small objects, objects of scale variation and objects with serious background interference. Class imbalances come from the number of different classes of objects and sampling of positive and negative samples. Due to these challenges, conventional object detection models usually cannot effectively detect objects in aerial images, especially in the balance between network speed and accuracy. In this paper, the YOLOv3 network structure was improved and an object detection method under the aerial visual scene (AVS-YOLO) was proposed. By introducing a type of densely connected feature pyramid strategy, a scale-aware attention module was constructed, considering both residual dense network blocks and the median-frequency-balancing mechanism. On this basis, an algorithm with ideal speed and accuracy for object detection is obtained. To verify the effectiveness of the algorithm, AVS-YOLO and YOLOv3 were both used to test the VisDrone-DET2019 and UAVDT. The experimental results show that the AP of AVS-YOLO increases by 6.22% and 5.09% on the VisDrone2019 and UAVDT datasets, respectively, compared with YOLOv3. In addition, the AP of AVS-YOLO is 1.82% higher than that of YOLOv4 on the VisDrone2019 dataset. In terms of detection speed, AVS-YOLO can process 31.8 frames per second on a single Nvidia GTX 2080Ti GPU, compared with 44.1 frames per second for YOLOv3. Compared with the other one-stage network in the field of object detection, AVS-YOLO currently achieves the state-of-the-art performance with similar calculation amount on this dataset.

There are many problems such as strong subjectivity, complex calculations, and lack of intelligence in most of the current energy system evaluation models, so a design of an evaluation system for urban smart energy systems based on an improved genetic algorithm-back propagation (BP) neural network is proposed. First of all, the hierarchical structure of the indicator evaluation system for energy system was established, analytic hierarchy process (AHP) was used to assign weights to each indicator, and data samples were classified. Then, SeqGAN was used to expand the data set, which solved the difficult problem of data acquisition. Finally, the genetic algorithm was used to determine the best initial values of the weights and thresholds of the BP neural network structure parameters designed for this research. The simulation experiment results showed that the method in this paper has higher classification accuracy than the traditional method, and comprehensive evaluation model proposed can effectively evaluate the urban smart energy system.

In this work, we implement an ARMv8 function and performance simulator based on gem5 infrastructure, which is the first open source ARMv8 simulator. All the ARMv8 A64 instructions other than SIMD are implemented using gem5 ISA description language. The ARMv8 simulator supports multiple CPU models, multiple memory systems, and McPAT power model.