3D reconstruction techniques have seen a surge of interest in panoramic depth estimation, its effectiveness heavily reliant on its omnidirectional spatial field of view. Panoramic RGB-D datasets are unfortunately scarce, stemming from a lack of dedicated panoramic RGB-D cameras, which subsequently restricts the practical implementation of supervised panoramic depth estimation techniques. Due to its reduced reliance on training datasets, self-supervised learning using RGB stereo image pairs holds the potential to overcome this limitation. Within this work, we detail the SPDET network, a self-supervised panoramic depth estimation architecture which integrates a transformer with spherical geometry features, emphasizing edge awareness. The panoramic geometry feature forms a cornerstone of our panoramic transformer's design, which yields high-quality depth maps. learn more Furthermore, a pre-filtering depth-image-based approach to rendering is employed to generate novel view images for the purposes of self-supervision. Meanwhile, a loss function attuned to image edges is being developed to enhance self-supervised depth estimation for panoramic images. In conclusion, we demonstrate the prowess of our SPDET via a suite of comparative and ablation experiments, reaching the pinnacle of self-supervised monocular panoramic depth estimation. Our models and code are located in the GitHub repository, accessible through the link https://github.com/zcq15/SPDET.
The emerging compression approach of generative data-free quantization quantizes deep neural networks to lower bit-widths independently of actual data. Full-precision network batch normalization (BN) statistics are instrumental in the data generation process by enabling network quantization. Although this is the case, there remains the consistent problem of decreased accuracy during application. Theoretically, we find that the variety of synthetic samples is integral for data-free quantization, but experimentally, existing methods, using synthetic data completely restricted by batch normalization statistics, show substantial homogenization problems at the distributional and individual sample levels. This paper's generic Diverse Sample Generation (DSG) scheme is presented to mitigate homogenization's detrimental effects in generative data-free quantization. Initially, the BN layer's features' statistical alignment is loosened to ease the distribution constraint. In the generative process, the loss impact of unique batch normalization (BN) layers is accentuated for each sample to diversify them from both statistical and spatial viewpoints, while minimizing correlations between samples. Our DSG's consistent performance in quantizing large-scale image classification tasks across diverse neural architectures is remarkable, especially in ultra-low bit-width scenarios. Data diversification, emerging from our DSG, improves the performance of various quantization-aware training and post-training quantization techniques, showcasing its broad applicability and effectiveness.
The Magnetic Resonance Image (MRI) denoising method presented in this paper utilizes nonlocal multidimensional low-rank tensor transformations (NLRT). We employ a non-local MRI denoising method, leveraging a non-local low-rank tensor recovery framework. learn more In addition, a multidimensional low-rank tensor constraint is utilized to obtain low-rank prior information, incorporating the 3-dimensional structural features of MRI image data. More detailed image information is retained by our NLRT, leading to noise reduction. The alternating direction method of multipliers (ADMM) algorithm is used to solve the optimization and update procedures of the model. Comparative experiments have been conducted on a selection of cutting-edge denoising methods. The performance of the denoising method was examined by introducing varying levels of Rician noise into the experiments and subsequently analyzing the obtained results. The results of our experiments confirm that our noise-reduction technique (NLTR) outperforms existing methods in removing noise from MRI scans, yielding superior image quality.
Medication combination prediction (MCP) serves to assist medical professionals in a more complete apprehension of the multifaceted processes involved in health and disease. learn more While recent studies commonly utilize patient representations from historical medical documents, the significance of medical understanding, encompassing prior knowledge and medication details, is often underestimated. A graph neural network (MK-GNN) model incorporating patient and medical knowledge representations is developed in this article, which leverages the interconnected nature of medical data. Precisely, patient features are extracted from their medical documentation, categorized into unique feature sub-spaces. After extraction, the features of patients are concatenated to create a unified representation. The relationship between medications and diagnoses, applied within pre-existing knowledge, generates heuristic medication features congruent with the diagnosis. Optimal parameter learning in MK-GNN models can be facilitated by these medicinal features. The medication connections in prescriptions are mapped to a drug network, merging medication knowledge with medication vector representations. Using various evaluation metrics, the results underscore the superior performance of the MK-GNN model relative to the state-of-the-art baselines. Through the case study, the MK-GNN model's practical applicability is revealed.
Certain cognitive research suggests that event segmentation in humans is a secondary outcome of event anticipation. Following this key discovery, we devise a simple yet effective end-to-end self-supervised learning framework for the delineation of events and the detection of their boundaries. Our framework, diverging from typical clustering-based methods, utilizes a transformer-based feature reconstruction approach for the purpose of detecting event boundaries via reconstruction errors. The ability of humans to discover new events is rooted in the difference between their predictions and the data they receive from their surroundings. The frames at the boundaries, due to their diverse semantic content, are hard to reconstruct accurately (frequently producing significant errors), which is beneficial for the identification of event boundaries. Consequently, given that reconstruction happens at the semantic feature level, not the pixel level, a temporal contrastive feature embedding (TCFE) module was designed to learn the semantic visual representation for frame feature reconstruction (FFR). This procedure, like human experience, functions by storing and utilizing long-term memory. The purpose of our work is to compartmentalize common events, as opposed to identifying specific localized ones. We strive to define the exact boundaries of each event with utmost accuracy. In conclusion, we employ the F1 score (precision in relation to recall) as our leading metric for a reasonable assessment in comparison with earlier strategies. Our calculations also include the conventional frame-based mean over frames (MoF) and the intersection over union (IoU) metric. Four publicly accessible datasets form the basis for our thorough benchmark, yielding much improved outcomes. Within the GitHub repository, https://github.com/wang3702/CoSeg, one will find the CoSeg source code.
Nonuniform running length, a significant concern in incomplete tracking control, is scrutinized in this article, focusing on its implications in industrial processes, particularly in the chemical engineering sector, and linked to artificial or environmental shifts. Iterative learning control (ILC), whose efficacy hinges on strict repetition, influences its application and design in critical ways. Accordingly, a dynamic neural network (NN) predictive compensation scheme is proposed within the context of point-to-point iterative learning control. For the purpose of tackling the complexities in establishing an accurate mechanism model for real-world process control, a data-driven approach is also utilized. Radial basis function neural networks (RBFNN) are integrated with the iterative dynamic linearization (IDL) technique to create an iterative dynamic predictive data model (IDPDM) predicated on input-output (I/O) signals. The model then defines extended variables to compensate for any incomplete operation duration. Through the application of an objective function, a learning algorithm relying on multiple iterative error measurements is presented. The NN continuously updates this learning gain to accommodate shifts within the system. The compression mapping, in conjunction with the composite energy function (CEF), underscores the system's convergence. In conclusion, a pair of numerical simulation examples are provided.
In graph classification, graph convolutional networks (GCNs) show superior performance, due to their structure which aligns with an encoder-decoder paradigm. Yet, most existing methodologies fail to adequately account for both global and local aspects during the decoding phase, causing the loss of global information or neglecting relevant local information in large-scale graphs. Essentially, the widely used cross-entropy loss is a global measure applied to the entire encoder-decoder system, neglecting to provide specific feedback on the training states of the encoder and decoder independently. We advocate for a multichannel convolutional decoding network (MCCD) as a solution to the problems discussed previously. A multi-channel graph convolutional network encoder is adopted first in MCCD, leading to superior generalization capabilities when contrasted with a single-channel GCN encoder. This is attributed to the differing perspectives offered by multiple channels in extracting graph information. Finally, we present a novel decoder that learns from global to local to decode graph information, subsequently resulting in better extraction of both global and local elements. To ensure the encoder and decoder are sufficiently trained, we implement a balanced regularization loss that supervises their training states. Our MCCD's efficacy, measured by accuracy, processing time, and computational cost, is demonstrated through experiments on standard datasets.