Training neural radiance fields without pre-estimated camera poses remains challenging, particularly in the presence of complex illumination and large inter-frame motion. This work presents a self-calibrating radiance-field framework designed to address these difficulties through a unified optimization strategy. Reliable monocular depth priors from the Depth Anything V2 model are incorporated to provide stable geometric constraints, and multiview point-cloud alignment is enforced to prevent pose drift under wide baselines. Illumination-induced photometric inconsistencies are handled by a jointly optimized exposure-compensation module that separates transient brightness variations from the underlying scene radiance. To enhance the representational capacity beyond that of standard multilayer perceptrons, a Kolmogorov-Arnold Network is adopted as the radiance-field backbone, enabling improved recovery of fine geometric and appearance details. The complete pipeline is deployed on a cloud-native infrastructure that supports scalable optimization and interactive access through an LLM-assisted interface. Experiments on indoor and outdoor datasets demonstrate that the proposed method consistently improves pose accuracy and novel-view synthesis quality over existing self-calibrating approaches.

Neural Radiance Fields (NeRF) have revolutionized 3D scene reconstruction through implicit neural representation, yet face fundamental limitations due to the C° continuity constraints and parameter inefficiency of traditional multilayer perceptrons (MLPs). This paper introduces NeRF-KAN, a novel approach that replaces MLPs with Kolmogorov-Arnold Networks (KANs) for radiance field modeling. Our key innovation lies in the dual-path KAN - Linear unit, which synergistically combines linear transformations for global structure capture with B-spline non-linear paths for local detail fitting, effectively resolving the global-local feature fragmentation inherent in MLPs. Additionally, we design a dual-term regularization strategy comprising activation and entropy regularization, specifically tailored to KAN's B-spline weight characteristics to mitigate overfitting in few-shot scenarios. Comprehensive experiments on both synthetic (Lego) and real-world (Fern) datasets demonstrate that NeRF-KAN achieves PSNR improvements of 0.5 dB (28.7→29.2 dB) and 0.8 dB (25.9→26.7 dB) respectively over original NeRF, while maintaining parameter efficiency. This work not only validates KAN's significant potential in 3D vision tasks but also establishes a new paradigm beyond traditional MLP-based approaches for neural radiance field modeling, with promising applications in cultural heritage digitization and medical imaging reconstruction.

To accurately measure the full-field deformation of soft-packed lithium-ion batteries and improve the safety of battery industrial production, an electronic speckle pattern interferometry (ESPI) system was established. After irradiating the object with a laser, uniform speckle patterns were formed on the camera target plane. Quantitative measurement of battery deformation was conducted via charging and discharging experiments on lithium iron phosphate (LFP) lithium-ion batteries. Through the automatic control of the reference surface phase shifting device and the industrial camera, four-step phase shifting was completed, and the wrapped phase was acquired. A phase unwrapping algorithm combining temporal and spatial domains was proposed to obtain the absolute phase variation, thereby achieving battery deformation measurement. The results show that the system exhibits excellent deformation measurement capability.

Clothes-Changing person re-identification algorithm (Re-ID) is the task of retrieving the query person in the case of the change of pedestrian clothing. Changes in pedestrian clothing lead to an offset in clothing features, resulting in a decrease in identification performance. Simply removing clothing may lead to the loss of contour information. Furthermore, algorithms based on feature decoupling cannot guarantee the accuracy of the positional relationship between clothing features and other features due to the lack of groundtruth. To address these issues, we propose a novel clothes-changing person re-identification algorithm based on causal feature supervision decoupling. Utilizing a multi-scale feature fusion module to extract fine-grained features of clothing and add supervised information labels. This enables the dual-branch network to separately approach overall and clothing features, promoting the extraction of effective identity information by the causal decoupling module, and obtaining unbiased estimations of pedestrians. Experimental results show that the proposed algorithm achieves the highest mAP and Top-1 accuracy on the LTCC and PRCC datasets. The source code is available at https://github.com/zhihu250/CISupNet.

The scarcity of high-quality question answering (QA) data remains a significant bottleneck in the development of intelligent educational systems, as traditional datasets lack the necessary scale and diversity to support personalized model training. To address this challenge, we propose the Book2QA framework, which integrates multiple language models to provide an effective and flexible approach for generating QA datasets derived from textbook content. To further enhance the quality of the generated data, we implement a hierarchical prompting strategy grounded in Bloom’s taxonomy, substantially increasing both the depth and breadth of the QA datasets. In addition, we fine-tune our model using these data, with evaluations conducted by both human reviewers and GPT-4 confirming its strong performance in real-world questioning scenarios. Experimental results demonstrate that our framework excels not only in specific textbook domains but also shows promise for broader applications across diverse fields. We open source our data and code at https://github.com/Curtain2020/Book2QA.

Knowledge tracing models have long grappled with the dual challenges of data sparsity and the limited ability to capture group learning patterns. Current contrastive learning paradigms in knowledge tracing (e.g., CL4KT framework) primarily employ sequence augmentation strategies to alleviate data scarcity constraints. However, such approaches frequently compromise semantic coherence during the augmentation process while failing to account for inherent similarity patterns within learner cohorts. To overcome these limitations, this paper introduces the GSCL-KT model (Group Similarity Contrastive Learning for Knowledge Tracing), which, for the first time, incorporates a group-similarity-aware contrastive learning mechanism into the knowledge tracing domain. Unlike traditional approaches that rely on manual data augmentation, GSCL-KT dynamically identifies positive and negative sample pairs from educationally homogeneous groups, enabling the discovery of group-level cognitive patterns while maintaining semantic coherence. The proposed model incorporates several advanced optimization strategies, including the Talking-Heads attention mechanism for fine-grained interaction modeling, the ContraNorm method for feature distribution regularization, and a correlation network enhanced by label dependencies. Experimental results on four real-world educational datasets demonstrate that GSCL-KT consistently outperforms existing baseline models, achieving the highest AUC and competitive performance across metrics.

ABSTRACT Automatic modulation recognition enables rapid spectrum access and serves as a key technical component for achieving communication‐aware integration. In recent years, deep learning methods have attracted considerable attention in this field. However, their performance relies heavily on the availability of large‐scale datasets. Data augmentation has proven effective in mitigating data scarcity. To address this challenge, this paper proposes a data augmentation algorithm based on a conditional diffusion model to improve model training under limited data conditions. In the proposed framework, the noise observation model detects noise in the input signal at different time steps. The identified noise is then iteratively removed during the reverse process of the conditional diffusion model to generate the corresponding modulated signals. Generated signals with high confidence are incorporated into the training set to enhance data diversity. Experimental results demonstrate that the proposed algorithm significantly improves the classification network's performance and outperforms existing data augmentation approaches.

In recent years, personality has been considered as a valuable personal factor being incorporated into the provision of personalized learning. Although some studies have endeavored to obtain learners’ personalities implicitly from their learning behaviors, they failed to achieve satisfactory prediction performance. On the one hand, most existing approaches ignore the imbalanced distribution of personality classes, which causes the personality classifiers to be biased toward the non-extreme personality class. On the other hand, the related methods normally focus on constructing statistical behavior features, while the sequence information of learning behaviors is ignored, but actually it can reflect learners’ behavior patterns more finely. In this paper, inspired by the human learning strategy in the face of small samples, we propose an effective Demographic-Guided Behavior Patterns Contrast (DGBPC) model to classify learners’ personalities through the demographic-guided contrast of learners’ coarse behavior patterns. Besides, we construct and publish the Personality and Learning Behavior Dataset (PLBD), which should be one of the largest public datasets regarding Big-Five personality and learning behavior sequence according to our knowledge. The experimental results on PLBD demonstrate that our DGBPC model could generate learner representations with higher discrimination and outperform the related methods in terms of balanced accuracy.

The goal of document-level relation extraction (RE) is to identify relations between entities that span multiple sentences. Recently, incomplete labeling in document-level RE has received increasing attention, and some studies have used methods such as positive-unlabeled learning to tackle this issue, but there is still a lot of room for improvement. Motivated by this, we propose a positive-augmentation and positive-mixup positive-unlabeled metric learning framework (P3M). Specifically, we formulate document-level RE as a metric learning problem. We aim to pull the distance closer between entity pair embedding and their corresponding relation embedding, while pushing it farther away from the none-class relation embedding. Additionally, we adapt the positive-unlabeled learning to this loss objective. In order to improve the generalizability of the model, we use dropout to augment positive samples and propose a positive-none-class mixup method. Extensive experiments show that P3M improves the F1 score by approximately 4-10 points in document-level RE with incomplete labeling, and achieves state-of-the-art results in fully labeled scenarios. Furthermore, P3M has also demonstrated robustness to prior estimation bias in incomplete labeled scenarios.

Children’s speech recognition plays an important role in the education research of children. The usual automatic speech recognition (ASR) systems are not satisfactory in terms of speech recognition for children, mainly due to the lack of child speech corpus. In recent years, there have been a large number of related studies on children’s speech recognition, most of which are improved based on the existing small amount of data through data augmentation, transfer learning, etc, which the promotion is limited. The purpose of our research is to establish a children’s speech corpus for children’s speech recognition. Firstly, we use the children’s language proficiency assessment system to collect weakly labeled children’s speech data from 3 to 12 years old, and obtain transcribed text through a commercial-grade speech recognition interface. Then, this paper proposes a method to quickly verify the transcribed text, and screens out effective label data pairs by combining children’s pronunciation and Chinese pinyin characteristics, and constructs a corpus of 50+ hours Mandarin children’s speech. We experimented with the current mainstream mandarin’s end-to-end pretrained model, finetuning on the data we built, and can achieve a 15% CER reduction on the basis of baseline without using finetuning.

In the Visual Question Answering (VQA) task context, most methods are influenced by language bias, resulting in poor performance on out-of-distribution data. Recently, some works attempted to use the adaptive margin loss to address this bias issue. However, these works typically consider only the frequency of answer labels when designing margin loss, leading to some samples being overly emphasized or lacking sufficient attention during model training. To address this issue, we propose a novel margin loss guided by the Gini-impurity for VQA debiasing. By comprehensively considering label distribution and instance complexity, we use Gini impurity to adjust the margin values in margin loss, balancing the attention of the model to different samples. Importantly, our method is plug-and-play and can be directly applied to any baseline. In the VQA-CP v2 task, our evaluation results across various baselines surpass the current state-of-the-art methods.

Multimodal abstract summarization is increasingly attracting attention due to its ability to synthesize information from different source modalities and generate high-quality text summaries. Concurrently, there has been significant development in multimodal abstract summarization models for videos. These models are capable of extracting information from multimodal data and generating abstract summaries. Most existing modeling approaches primarily concentrate on instructional videos, such as those teaching sports or life skills, thereby limiting their ability to capture the complexity of dynamic environments in the general world. In this paper, we propose a vision-guided model for multimodal abstractive summarization with knowledge distillation KD-VSUM to address the lack of generalized video domain capabilities in video summarization. This approach includes a vision-guided encoder, which enables the model to better focus on the global spatial and temporal information of video frames. We capitalize on knowledge distillation from multimodal pre-trained video-language models to enhance model performance. We introduce the VersaVision dataset, which includes a broader range of video domains and a higher proportion of medium to long videos. The results demonstrate that our model surpasses existing state-of-the-art models on the VersaVision dataset, achieving ROUGE scores of 1.7 in ROUGE-1, 1.8 in ROUGE-2, and 2 in ROUGE-L. These findings underscore the substantial improvements that the integration of a global vision guided and knowledge distillation can bring to the task of video summary extraction.

Abdominal lymph nodes (ALNs) play a pivotal role in tumor metastasis, necessitating their early detection for tumor diagnosis and treatment decisions. However, due to varying ALN sizes and complex abdominal environment, existing methods face challenges, including small ALN absence, high False Positive (FP) rate and low efficiency. This study aims to propose a two-stage ALN detection approach based on multi-scale 2.5D input to overcome these difficulties. It leverages multi-scale images and the Skip-Attention module to generate ALN cadidate regions in the first stage. Then, a new 2.5D input is constructed by extracting slices from different planes and identifing genuine ALNs via the FP reduction algorithm. We validated our method with the dataset containing both public and real patient data. The first stage achieved an 81.1% detection rate with 21 FP/case, while the second stage reached an AUC value of 0.94 after ensemble learning. The final cascade experiment illustrates that our model achieved a 73.5% detection rate, surpassing other state-of-the-art methods including 3D models by at least a margin of 3.3% and significantly improving time efficiency.

The effectiveness of immune checkpoint blockade (ICB) therapy, a critical strategy in cancer immunotherapy, is often limited by a high rate of primary resistance. Identifying patients likely to respond to this therapy is a significant challenge. To tackle this issue, we introduce a novel deep learning model known as the Permutable Hybrid Zipped Network (PHZNet), which is specifically tailored for classifying patient responses to Immune Checkpoint Blockade (ICB) therapy. We achieved 89.67% accuracy on the transcriptomics public dataset from the Faculty of Pharmacy, Universit'e de Montr'eal, outperforming other excellent models. In addition, to explore interpretability, based on the concept of model distillation, we propose to use XGBoost gradient boosting decision trees to distill PHZNet and thus uncovering key gene features influencing therapy response. Our study provides a reliable basis for clinical decision-making by predicting whether patient benefits from ICB therapy across pan-cancer. Furthermore, it paves the way for understanding the ICB treatment response mechanism and provides critical gene targets for future immune treatment strategies.

Document-level relation extraction (RE) aims to identify relations between entities across multiple sentences. Most previous methods focused on document-level RE under full supervision.However, in real-world scenario, it is expensive and difficult to completely label all relations in a document because the number of entity pairs in document-level RE grows quadratically with the number of entities.To solve the common incomplete labeling problem, we propose a unified positive-unlabeled learning frameworkshift and squared ranking loss positive-unlabeled (SSR-PU) learning.We use positive-unlabeled (PU) learning on documentlevel RE for the first time.Considering that labeled data of a dataset may lead to prior shift of unlabeled data, we introduce a PU learning under prior shift of training data.Also, using none-class score as an adaptive threshold, we propose squared ranking loss and prove its Bayesian consistency with multi-label ranking metrics.Extensive experiments demonstrate that our method achieves an improvement of about 14 F1 points relative to the previous baseline with incomplete labeling.In addition, it outperforms previous state-of-the-art results under both fully supervised and extremely unlabeled settings as well. 1

Studies on recommender systems persuaded accuracy of predictions, leading to the filter bubble problem which narrowed the user's interest areas. Unexpected recommendation (UR) is one of the solutions in handling the filter bubble. However, existing research on UR still faces two shortcomings. On the one hand, it is difficult to measure the unexpectedness of items to users accurately. On the other hand, there is a decrease in accuracy when making unexpected item recommendations. In this paper, we propose an Unexpected Deep Interest Network (Unexp-DIN) model to solve these problems. We use mean-shift for multiple clustering and Cluster-attention to maximize the unexpectedness. Then we use the constraint function and the hybrid utility function to simultaneously improve accuracy and unexpectedness. Experiments on three publicly available benchmark datasets show that our model has different degrees of improvement in both accuracy and unexpectedness.

With the application of artificial intelligence technology in the medical field, the use of deep learning methods to predict lymph node (LN) metastasis from Computed Tomography (CT) images has become one of the important studies in adjuvant cancer treatment. Although many studies have made some progress, the large difference in LNs size has always limited the performance of traditional convolutional neural network methods. In this paper, we propose a multi-scale attention fully convolutional network for LN metastasis prediction in CT images. We aim to make the network compatible with LNs of different sizes to improve the overall performance of the network. First, the network extracts image features from multiscale input data. Then, we use an attention-based fusion module to study the relationship between features at different scales and adaptively fuse image features. Furthermore, a feature consistency loss is proposed by us to enhance the similarity between image features at different scales. The experimental results show that our proposed network achieves the best performance with the accuracy 92.37%, the sensitivity 90.21% and the specificity 95.71%, which outperforms several state-of the-art methods.

In recent years, session-based recommendation has attracted more and more attention. Previous work utilizing deep learning approaches has achieved significant progress in the accuracy of prediction. However, why the user would view the next item is not clear, though attention mechanism assigns different importance on items in a session. Meanwhile, some traditional data mining-based methods are able to provide explanations, but they mainly offer a narrow perspective like measuring item similarity and fail to achieve as good performance as deep learning-based approaches. To this end, we propose an explainable session-based recommendation by considering three factors including sequential patterns, repetition clicks, and item similarity. Concretely, we design a two-stage framework where candidate items are firstly selected according to users' sequential patterns and repeated behaviors, and then they are further ranked by considering short-term interest with the calculation of item similarity. The experimental results on two benchmark datasets show that our model not only achieves competitive recommendation performance with advanced deep learning based models, but also provides reasonable explanations.

Jointly extracting entities and classifying relations aims to detect all possible triples from unstructured text with a single model.Tagging-based method effectively improves the performance of jointly relation extraction.However, some taggingbased approaches ignored that one entity pair may exist multiple relations and others set an empirical threshold value for selecting one or more relevant relations, which becomes the bottlenecks of the model.As a solution, we propose the attention guided filter, namely, AGFRel, which introduces transformer blocks to learn the number of relations for every entity pair to filter out irrelevant relations.Moreover, each module of the model has a multi-head attention guided layer to highlight valuable information.Extensive experimental results show that AGFRel is capable of gaining better performance on various tasks including overlapping triples extraction and multiple triples extraction.On NYT and WebNLG public datasets, our model obtains F1 score 90.8 and 91.9 respectively and achieves a new state-of-the-art performance.

In recent years, Time-sync Comment (TSC), as known as “Danmaku” or “Danmu”, has been increasingly recognized as a valuable representation of video being incorporated into the process of generating video or highlight recommendation. However, little work has studied how to recommend proper TSC when users are watching the video. Providing some suggestions for users when they want to post TSCs can not only enhance the real-time interactions among users within videos, but also motivate users to post more TSCs which could be useful for video or highlight recommendation in return. In order to accomplish the TSC recommendation task, we extract candidates from existing comments sent by other users. Specifically, we propose a model, namely “TSCREC”, which uses a bidirectional Gated Recurrent Unit to capture the semantic meaning of existing comments and assigns scores for comments by a Multi-Layer Perceptron. Furthermore, considering the importance of correlations among comments, we also take similarities among comments as features. In addition, we design a set of novel evaluation metrics which combines n-grams overlapping and ranking order to measure the quality of the recommendation list. We conduct experiments on realworld datasets, and the results show that our TSCREC model outperforms baseline approaches.

Crowd counting, i.e., estimation number of the pedestrian in crowd images, is emerging as an essential research problem with the public security applications. The density-based method of crowd counting still has some challenges, such as lack of perspective information in density map and background noise. Current models often misjudge background noise as a person and the ground truth density map widely used now is not so accurate. In this paper, we present a novel approach to help generate a higher quality density map. On the one hand, we eliminate the apparent mistakes in the density map with the help of a semantic segmentation model, which provides more information about fine-granted negative samples. On the other hand, we modify the density map to make sure it maintains a natural attribute. The experimental results prove the effectiveness of our method for crowd counting models, especially in uneven distribution monitoring scenario.

Crowd counting, i.e., estimation number of the pedestrian in crowd images, is emerging as an important research problem with the public security applications. A key component for the crowd counting systems is the construction of counting models which are robust to various scenarios under facts such as camera perspective and physical barriers. In this paper, we present an adaptive scenario discovery framework for crowd counting. The system is structured with two parallel pathways that are trained with different sizes of the receptive field to represent different scales and crowd densities. After ensuring that these components are present in the proper geometric configuration, a third branch is designed to adaptively recalibrate the pathway-wise responses by discovering and modeling the dynamic scenarios implicitly. Our system is able to represent highly variable crowd images and achieves state-of-the-art results in two challenging benchmarks.

The accuracy of OCR is usually affected by the quality of the input document image and different kinds of marred document images hamper the OCR results. Among these scenarios, the low-resolution image is a common and challenging case. In this paper, we propose the cascaded networks for document image super-resolution. Our model is composed by the Detail-Preserving Networks with small magnification. The loss function with perceptual terms is designed to simultaneously preserve the original patterns and enhance the edge of the characters. These networks are trained with the same architecture and different parameters and then assembled into a pipeline model with a larger magnification. The low-resolution images can upscale gradually by passing through each Detail-Preserving Network until the final high-resolution images. Through extensive experiments on two scanning document image datasets, we demonstrate that the proposed approach outperforms recent state-of-the-art image super-resolution methods, and combining it with standard OCR system lead to signification improvements on the recognition results.

Electronic Medical Records (EMRs) are the digital equivalent of paper records, which include treatment and medical history about a patient. At present, the main research goal of Chinese EMRS is to accurately recognize the body parts, drugs, illnesses and other information in the Chinese medical process. Implementing EMRs can boost both the quality and safety of patient care. In Chinese EMRs, how to accurately recognize named entities is important because it is useful to predict the disease risk, therapeutic method and recovery probability. This paper proposes a novel deep learning framework, which uses character-word joint embedding and combines different feature information based on the dictionary. Compared with the predecessors, we incorporate word-level information based on the basic Bi-LSTM model. In addition, we propose an improved n-gram feature encoding method and compare it with PDET feature and PIET feature. Our experimental results demonstrate that our proposed model performs the best in predicting named entities in Chinese EMRs.