Recently, personalized federated learning (pFL) has attracted increasing attention in privacy protection, collaborative learning, and tackling statistical heterogeneity among clients, e.g., hospitals, mobile smartphones, etc. Most existing pFL methods focus on exploiting the global information and personalized information in the client-level model parameters while neglecting that data is the source of these two kinds of information. To address this, we propose the Federated Conditional Policy (FedCP) method, which generates a conditional policy for each sample to separate the global information and personalized information in its features and then processes them by a global head and a personalized head, respectively. FedCP is more fine-grained to consider personalization in a sample-specific manner than existing pFL methods. Extensive experiments in computer vision and natural language processing domains show that FedCP outperforms eleven state-of-the-art methods by up to 6.69%. Furthermore, FedCP maintains its superiority when some clients accidentally drop out, which frequently happens in mobile settings. Our code is public at https://github.com/TsingZ0/FedCP.

With the increase in data volume and environment complexity, real-world problems require more advanced algorithms to acquire useful information for further analysis or decision-making. Cognitive Learning effectively handles incomplete information, and multi-agent systems can provide enough data for analysis. Inspired by Distributed Machine Learning, Federated Learning has become an efficient framework for implementing Cognitive Learning algorithms in multi-agent systems while preserving user privacy. However, traditional communication optimizations on the federated learning framework suffer from either large communication volumes or large accuracy loss. In this paper, we propose pFedEff, a personalized federated learning framework with efficient communication that can reduce communication volume and preserve training accuracy. pFedEff uses two magnitude masks, two importance masks, and a personalized aggregation method to reduce the model and update size while maintaining the training accuracy. Specifically, we use a pre-training magnitude mask for approximated regularization to reduce the magnitude of ineffective parameters during training. We also use a post-training magnitude mask to eliminate the low-magnitude parameters after training. Furthermore, we use uploading and downloading importance masks to reduce the communication volume in both upload and download streams. Our experimental results show that pFedEff reduces up to 94% communication volume with only a 1% accuracy loss over other state-of-the-art federated learning algorithms. In addition, we conduct multiple ablation studies to evaluate the influence of hyperparameters in pFedEff, which shows the flexibility of pFedEff and its applicability in different scenarios.

The most methods of close-range photogrammetry are based on Direct Linear Transformation (DLT). But DLT often has unstable solution and every image needs more than six ground control points to compute DLT parameters, so this method is hard to acquire the high accuracy and its efficiency is low. The paper discusses a new method of digital close-range photogrammetry - panning and multi-baseline digital close-range photogrammetry. This method enlarges the intersection angle and improves the intersection precision by multi-baseline. At the same time this method applies the classic aerotriangulation and bundle adjustment to the close-range photogrammetry, we need more than three ground control points to compute the exterior orientation elements of all images. The experiments prove that this method can acquire the high accuracy.

Multiple range images registration is one of the most important problems in range image analysis. This paper proposes a new method to implement seamless registration of multiple range images under least squares, since all the range images constitute a circled network, which can be seen as the closed condition of the whole block adjustment. We experimented on real range images taken by laser scanners, and observed that our method worked successfully even for noise data, and it reduces significantly the level of the registration errors between all pairs in a set of range images. The proposed method has the distinct advantage of seamless and robustness