Graph neural networks (GNNs) have proven to be powerful tools for learning from graph-structured data and have achieved great success in many applications. As the sizes of real-world graphs continue to grow, traditional GNN training methods face significant scalability challenges. Recently, disks have gained attention as a cost-effective solution to store large-scale graphs, and several disk-based GNN systems have been proposed to train large-scale graphs on a single machine. However, these systems either overlook the unique data characteristics of GNN workloads when designing cache plans or fail to fully exploit the multilevel hierarchy of storage and computation in system execution, thus resulting in disk I/O bottleneck and resource under-utilization. To address these issues, we present CaliEX, an advanced disk-based GNN system that employs joint optimizations of caching and execution within and across different training stages. CaliEX first designs tailored cache plans and execution policy for both graph topology and features to accelerate neighborhood sampling and feature gathering. Since these two training stages work on different types of data, CaliEX further auto-tunes the cache allocation and pipelines the execution across different stages to improve resource utilization and overall training throughput. Evaluations on multiple GNN models and various large-scale datasets show that CaliEX achieves 3.28 × speedup on average compared to existing disk-based GNN training systems.

Existing person re-identification (ReID) methods mainly rely on images and videos to match persons across cameras, yet visual data captured by cameras are vulnerable to environmental interferences (e.g. illumination and occlusion) or personal appearance changes, leading to performance degradation under such scenes. Meanwhile, the popularization of Wi-Fi networks has allowed probe requests to be captured for mobile sensing applications such as crowd counting and trajectory estimation. However, the MAC address randomization technique adopted by modern devices breaks the association of probe requests and adversely affects the functionality of these applications. In this paper, we propose MaRPA, the first multimodal training approach that incorporates both videos and Wi-Fi probe requests to simultaneously promote tasks of probe requests association and person ReID. MaRPA first distinguishes among pairwise probe request frames through a contrastive learning model. It then matches video and probe request sequences by exploring their similarities from the position and the vision aspects. Matched videos and probe requests provide complementary information and generate more robust features for both tasks. To evaluate MaRPA, we contribute a new dataset containing synchronous videos and probe requests data for probe requests association and person ReID. Experimental results demonstrate the effectiveness of our approach. For probe requests association, it achieves > 85% discrimination accuracy and > 0.90 V-measure score; for person ReID, it achieves 75.8% mAP and 90.6% Rank-1, improving state-of-the-art video-based ReID methods by over 40%

Scene matching involves establishing correspondences between multiple images of the same location and poses significant challenges for synthetic aperture radar (SAR) images due to the anisotropic scattering prosperities of SAR targets; variations in looking and azimuth angles further complicate the matching process. A matching algorithm is proposed based on a coarse-to-fine framework to address these issues. First, a coarse matching employing normalized cross correlation (NCC) with a sliding window is applied to filter out irrelevant regions, reducing distractions, and shortening the processing time. Subsequently, a Siamese neural network (SNN), incorporating ResNet-50 and convolutional block attention module (CBAM) for enhanced feature extraction, is introduced to learn and discern differences between inputs. The effectiveness and robustness of the proposed method are validated through extensive experiments using a self-made dataset derived from Umbra Satellite.

Deep learning-based SAR image interpretation have gained much attention in recent years, in this paper supervised learning for SAR images is discussed first and then novel self supervised contrastive learning method is presented to reduce dependency on large amount of labeled data. SAR images provide valuable information about earth surface. Unlike optical images, SAR images are formed by backscattered signals influenced by surface roughness and other physical properties. Deep learning models can extract details from SAR data that may be missed by traditional image processing. Furthermore, deep learning models can learn patterens and features from SAR images to enable automated interpretation and analysis. It is difficult and time consuming to get labeled data. Supervised CNN is used to extract meaningful data from the SAR images for classification but it uses labeled data extensively, using self-supervised contrastive learning, positive pairs are constructed for each image; positive pairs are generated using augmentation of the SAR image. The contrastive loss function is designed such that it encourages the network to learn similar representation for positive pairs. A novel features extractor is introduced to get robust features. Contrastive loss and backpropagation are used to train the network without labels, after training, fine tuning is done on small number of labels called few shot. Downstream tasks such as classification is performed after fine tuning. Results are evaluated on the widely used SAR bench mark dataset.

Ship target detection based on SAR images is an important means of marine observation. Traditional target detection requires the most processing time to image the SAR echo. Considering the sparse distribution of ship targets in wide-swath marine SAR images, imaging and detecting processes on non-target regions seriously reduce efficiency. This paper proposes an integrated framework to improve marine SAR imaging detection efficiency by adding two steps of selection for target areas. Firstly, an RC-TextCNN network is designed to select target areas on azimuth direction from SAR echo one-dimensional compression data. After imaging selected areas, a dynamic quantization and threshold segmentation method is used to further remove non-target areas. Finally, suspected target areas are introduced into the pruned yolov7 model for final target detection. This workflow significantly minimizes computational and time costs. The experiment on Gaofen3 data shows that the speed of the process is increased by three times while detection accuracy is at 90%.