A dynamic (or temporal) network is a widely used structure that enables understanding dynamic systems by modeling interactions among system components over time. In many real-world cases, however, components (called nodes) and/or interactions (called edges) contain numerous meaningful attributes, leading to the need for a more suitable instrument for representing and analyzing these dynamic and complex systems with multiple attributes: the Dynamic Multivariate Network (DMVN). In this work, we extended LargeNetVis, a visualization system specifically designed for large dynamic networks that focus on network community structure and dynamics, to enable the visual exploration of DMVNs and their communities. The newly introduced visual encodings and interactions allow the visualization of nodes' and edges' attributes at different granularity levels and produce a node tracking capability from both top-down and bottom-up perspectives. With these functionalities, one can track individual nodes across dynamic communities over time. The proposed approach is validated by comparing it with the original LargeNetVis system and conducting a user evaluation involving 37 participants.
Building Information Modeling (BIM) integrates engineering data with 3D geometry, spatial relationships, and other properties to support construction projects throughout their entire lifecycle. Researchers have been focusing on how to automatically generate or update these models from 3D point clouds, considering existing buildings. Most of these studies have focused on recognizing planar structures (such as floors and walls) or cylindrical ones (such as pipelines). Only a few recent works have addressed the detection of steel structural elements, due to their particular geometry. In these approaches, the point cloud of each structural element was manually separated from the point cloud of the entire building. This practice poses a challenge, as manual segmentation of point clouds is a timeconsuming and subjective process. In this paper, we propose a new approach to automatically detect structural steel elements in 3D point clouds of existing buildings projects, without the need for prior segmentation and regardless of orientation. The proposed technique combines geometry processing algorithms with a machine learning strategy. The performance results demonstrate the effectiveness of our approach in detecting the desired elements. These results open up new possibilities for developing automated pipelines aimed at generating highly accurate 3D BIM models.
Generating synthetic data with diffusion models poses a challenge due to the scarcity of data in computer vision. However, its practical application presents challenges, including the need for computationally expensive retraining and the complex, manual effort required to design precise textual instructions. This work approaches these limitations by introducing the ZEDD-G method, a novel zero-shot and prompt-free framework for synthetic data augmentation specifically designed to enhance downstream classification tasks. Our methodology establishes a fully automated pipeline that begins with unsupervised clustering of visually similar images, followed by a multi-image latent guidance mechanism. This mechanism combines visual prompts with direct manipulation of latent features to generate diverse and controlled variations. Evaluated on a demanding classification benchmark, which includes natural and medical images, ZEDD-G demonstrates a substantial impact on performance. For a ResNet-50 model trained from scratch, our method boosts accuracy by an average of over 30 percentage points across diverse datasets. For a pre-trained model, it provides consistent gains of around 2.6 percentage points. These results are competitive with state-of-the-art fine-tuning-based methods, achieved without requiring any model retraining. Thus, ZEDD-G is an efficient method that enables the generation of high-quality synthetic data, thereby improving classification tasks. Our implementation is publicly available at https://github.com/Gardiy/ZEDDG.
Glacial lake outburst floods (GLOFs) pose a significant threat to high-mountain communities and infrastructure, particularly in glacierized regions such as the Cordillera Blanca, Peru. This study presents an automated method for mapping glacial lakes using Sentinel-2 satellite imagery with enhanced band combinations and a segmentation-based foundational model (SAM 2.1). Our method enables consistent multitemporal lake extraction, which successfully mapped 80% of 448 manually identified glacial lakes. We performed a comparative analysis of images taken from May 2016 and May 2024, and identified five lakes potentially vulnerable to GLOFs. Notably, Lake Parón and Lake Piticocha experienced significant surface area expansion of 13.79 and 3.72 hectares, respectively. The other three lakes showed significally expansion despite its smaller size. GLOF simulations incorporating local topography and lake size indicate potential impacts on urban areas (e.g., Caraz city) and agricultural land. These findings highlight the importance of standarized and automated glacial lake monitoring for early risk assessment and disaster preparedness in the face of climate-driven glacial change.
The rapid development of Multimodal Large Language Models (MLLMs) has expanded the possibilities for automatically understanding complex real-world scenarios. However, their ability to interpret social interactions through videos remains underexplored. This study introduces BeSIM, a benchmark designed to evaluate MLLMs' competence in interpreting such interactions among Brazilians, based on the APRACE taxonomy for categorizing key elements of social interaction. A set of 22 videos was collected from YouTube, resulting in 110 multiple-choice questions aligned with these categories. The results show that models such as Gemini 2.5 Pro outperform their performance on generalist benchmarks like Video-MME, reaching up to 90 percent of accuracy on BeSIM. We also conducted a qualitative analysis to discuss possible features to explain failure/success of the models. These findings indicate that, when properly evaluated, MLLMs demonstrate great potential in interpreting human interactions. Code and data available at https://github.com/M4Falcao/BeSIM.
High-resolution volumetric imaging techniques, such as X-ray tomography and advanced microscopy, generate increasingly large datasets that challenge existing tools for efficient processing, segmentation, and interactive exploration. This work introduces new capabilities to Annotat3D through Harpia, a new CUDA-based processing library designed to support scalable, interactive segmentation workflows for large 3D datasets in highperformance computing (HPC) and remote-access environments. Harpia features strict memory control, native chunked execution, and a suite of GPU-accelerated filtering, annotation, and quantification tools, enabling reliable operation on datasets exceeding single-GPU memory capacity. Experimental results demonstrate significant improvements in processing speed, memory efficiency, and scalability compared to widely used frameworks such as NVIDIA cuCIM and scikit-image. The system's interactive, human-in-the-loop interface, combined with efficient GPU resource management, makes it particularly suitable for collaborative scientific imaging workflows in shared HPC infrastructures.
The connection between urban forests and phenomena such as heat islands, air pollution, property values, and public perceptions of safety has been the subject of extensive research over the years. A fundamental requirement for such studies is the accurate mapping of the spatial distribution of green areas. In this context, this work presents a concise and reproducible framework to map urban forests and generate datasets detailing their location and distribution within metropolitan regions. The proposed methodology relies on freely available imagery from the CBERS-4A satellite and the open-source QGIS software. While auxiliary layers such as city layout and shadow masks are used during the modeling process, their purpose is solely to enhance the accuracy of urban forest classification, not to serve as final outputs. The true merit of this approach lies in its ability to identify urban tree cover patterns with high accuracy, making it a scalable solution for urban green infrastructure studies. The framework is applied to map urban forests in the central area of the city of São Paulo, Brazil.
With the proliferation of image capture devices and social media, face recognition technology has become widespread, raising significant privacy concerns and leading to the development of data protection regulations like GDPR and LGPD. Reversible face anonymization techniques, which aim to obscure an individual's identity while allowing for its recovery with a secret key, have emerged as a promising solution to balance privacy and data utility. These methods, often based on Generative Adversarial Networks (GANs), generate realistic, anonymized faces that preserve the original image's structure. This paper challenges the presumed security of state-of-the-art reversible face anonymization systems. We introduce a novel attack that bypasses the password-based protection by exploiting information leaked in an intermediate stage of the anonymization pipeline. Our attack on a leading reversible anonymization method, RiDDLE, successfully recovers the original identity from anonymized images in up to 70.2 % of cases on the VggFace2 dataset, a stark contrast to the 0.2 % chance of a random guess. To counteract this critical vulnerability, we propose a new loss function, the Secure Image Embedding Loss, designed to be integrated into the training of anonymization models. This addition enforces the disassociation of the original identity from the intermediate representations. Experimental results demonstrate the effectiveness of our defense, reducing the success rate of our attack to a mere 0.8 % on VggFace2 and showing robust performance on the LFW dataset. These findings highlight a significant, previously overlooked security flaw in reversible face anonymization and provide a concrete solution to fortify these methods against such attacks, marking a crucial step toward ensuring genuine privacy in facial data.
Quantitative Wood Anatomy (QWA) is critical for understanding plant hydraulic function, ecological strategies, and environmental responses through the analysis of xylem anatomical traits. However, the advancement of automated image analysis in QWA has been hindered by the lack of large-scale annotated datasets, limiting the effective use of deep learning (DL) techniques. Existing datasets are small, domain-specific, and lack the diversity required for robust model generalization. This work addresses these limitations by introducing a large-scale, diverse annotated dataset derived from the Inside Wood repository. To establish an optimal annotation pipeline, we evaluated multiple semantic segmentation methods, which demonstrated that they trained on our dataset achieved a mean Intersection over Union (mIoU) exceeding 90 %, significantly outperforming models from scratch, pre-trained on ImageNet and fine-tuned on target-domain data for a real-world scenario. This results validate our dataset and annotation methodology as a strong foundation for developing accurate and generalizable segmentation models for practical and industrial QWA applications.
This paper presents a method for rendering static reflections using precomputed 3D Gaussian Splatting (3DGS) models, aiming to enhance conventional rendering pipelines with realistic and efficient reflections. Unlike traditional techniques such Environment Mapping, which suffer from occlusion limitations and static viewpoint assumptions, our approach decouples reflection generation from the main rendering process by sampling color and radiance from 3DGS models organized in an octree structure. By constraining the number of Gaussians per node, our method allows for scalable control over quality and performance, achieving real-time frame rates in several synthetic and real-world scenarios. We show that even with simplified models, the reflections remain visually plausible, particularly in scenes with diffuse materials. Our results demonstrate accurate parallax and off-camera content reproduction, with potential applications in games, mixed reality, and offline rendering.
Classification is a fundamental tool to automate the process of categorizing documents in many real-world applications, such as information management, financial document processing, healthcare records management, news categorization, fraud detection, regulatory compliance, and many others. Because of this broad spectrum of applications, document classification is of paramount importance for various companies. However, documents often change in terms of format and their visual patterns, which may impair a simple classification model. Moreover, model continuance and retraining often demands important efforts, consuming computational resources and demanding new data. Therefore, techniques capable of classifying documents by simply observing new data, without necessarily requiring retraining the classifier, are of immense importance for a wide variety of applications. In this context, Zero-Shot Learning (ZSL) is especially suitable for document classification because it handles diverse and ever-changing document content. In this work, we tackle the gap involving Zero-Shot Document Image Classification (ZS-DIC), where we classify documents that have not been seen by the model during training. To achieve this, we built Layout-Aware Complex Document Information Processing (LA-CDIP), a dataset tailored for this problem. LAC-DIP prioritizes structural consistency, allowing models to classify documents correctly under a ZSL scenario. To benchmark this dataset, we developed a series of Siamese Neural Networks (NNs) based on a variety of computer vision neural architectures, such as ResNet, EfficientNet, ViT and others. As a result, the proposed ZSL-based method achieves Equal Error Rates (EERs) under 5%. The code of the proposed method is available at https://github.com/ABMHub/doc-zsl.
Semi-supervised learning (SSL) is an effective approach to addressing the scarcity of accurately labeled data, leveraging both labeled and unlabeled samples for deep model training. However, limited labeled data can constrain its effectiveness. Active learning (AL) mitigates this by selecting informative samples for human annotation, enhancing SSL performance. Most state-of-the-art AL and SSL methods depend on pre-trained features and large validation sets for learning representations in classification tasks. We introduce contrastive active Deep Feature Annotation (ca-DeepFA), a method that integrates contrastive learning, active learning, and curriculum-based meta-pseudolabeling to train non-pre-trained CNNs for image classification with minimal labeled data and abundant unlabeled samples. The process begins with unsupervised contrastive pre-training on a small labeled set. At regular epoch intervals, label propagation is applied to 2D deep feature projections. Following this, pseudolabels are selected under a curriculum-driven policy, while an oracle annotates the most informative samples. These contribute to a composite loss function that combines supervised contrastive, supervised, and semi-supervised components-leading to enhanced feature representations for image classification. Evaluated on three real-world biological image datasets with a limited amount of labeled data, our method consistently outperforms baselines and state-of-the-art approaches by improving generalization and reducing annotation effort.
Protoplanetary disks (PPDs) are structures composed of gas and dust surrounding young stars, which play a fundamental role in understanding the formation of planetary systems. Despite their significance in astronomy, these objects remain underexplored in the image recognition literature. This study presents a comparative analysis of deep learning architectures for PPD classification, based on a catalog curated by domain experts. The dataset includes images captured by the Wide Field Channel of the Advanced Camera for Surveys (ACS/WFC) from the Hubble Space Telescope (HST) as part of the Orion Nebula Treasury program, along with annotations from the consolidated PPD catalog. To address class imbalance we rearrange the original classes and use data augmentation techniques of geometric image transformations and band combination. The experimental setup explores a range of neural network architectures and hyperparameters, including traditional models, such as VGG, ResNet, Inception and EfficientNet, as well as a more recent ConvNeXt architecture. An ablation study is also conducted to assess the contribution of NLM+Anscombe as a denoising step. The results indicate a general difficulty in learning discriminative patterns from the data. Among the evaluated architectures, EfficientNet achieved the best overall performance, with a mean F1-score of 53%, reaching a maximum of 65% using a learning rate of 0.02 and a weight decay of 0.0004. The denoising filtering showed no statistically significant impact on model performance for the specific configuration.
Low-light image enhancement (LLIE) is an important task in image processing and computer vision, addressing the need to improve the visual quality of images captured in suboptimal lighting conditions. Enhanced images are not only more visually appealing but also more effective for downstream tasks such as object detection and classification. However, while the desired level of enhancement is often subjective and varies across users and applications, previous LLIE techniques do not allow users to control the desired level of enhancement intensity. We introduce the Restoration Level Estimator (RLE) block, a novel component designed to provide control over the enhancement level in existing LLIE models. The RLE block can be seamlessly integrated into convolutional neural networks, adding a new channel that allows users to adjust the level of enhancement applied to input images. Our experiments show that, in addition to offering control, the RLE block can improve the overall performance of LLIE models as measured by PSNR, SSIM, and LPIPS. We demonstrate the flexibility of our approach across multiple LLIE models, highlighting its potential to improve both user experience and model performance.
Identifying cattle brands is a challenging visual task due to variability in branding styles, degradation over time, and the absence of standardized, large-scale datasets. To address this, we propose a deep learning-based Content-Based Image Retrieval (CBIR) framework specifically designed for this problem. Our system matches hand-drawn or digitized query sketches of cattle brands against a reference database of binarized brand symbols using learned visual embeddings for similarity search. To support training and rigorous evaluation, we assembled the Apporteira Cattle Brand Dataset, comprising 5,233 clean binary brand templates, 1,454 hand-drawn sketches, and over 627,000 augmented images simulating real-world distortions via rotations, morphological operations, and homographic transformations. This dataset has been made publicly available to enable reproducible research and benchmarking. We evaluate classical feature descriptors, pretrained convolutional neural networks (ResNet-50, MobileNet, VGG), and a fine-tuned VGG-16 model adapted to this domain. Experiments are reported with standard CBIR metrics, including mean Average Precision (mAP) and Top-$k$ accuracy, together with retrieval efficiency using Facebook AI Similarity Search (FAISS) to assess scalability. Our fine-tuned model achieves 79.71% Top-1 and 97.18% Top-10 accuracy, substantially outperforming generic CNN baselines and handcrafted methods. The results highlight the effectiveness of task-specific fine-tuning, showing consistent gains even when baselines perform strongly at higher ranks, and demonstrate the system's robustness to symbol variation, offering a scalable solution for livestock identification, rural security enforcement, and brand registry automation.
Understanding spatiotemporal anomalies is critical in domains such as urban safety, mobility, and environmental monitoring. These scenarios involve complex dynamics that are effectively modeled using graph-based representations, where the spatial structure is encoded through data connectivity, and each node corresponds to a time series. Anomaly detection in such data is crucial for identifying unusual or significant events, but it requires complex methods involving pattern recognition, prediction, and classification. Interpreting these anomalies remains challenging. To address this, we introduce an interactive system that combines spatiotemporal visualizations with Large Language Models (LLMs) to generate context-aware explanations by unifying temporal, spatial, and textual insights. We guide the LLM using a structured prompting strategy grounded in the data to reduce hallucinations and improve plausibility. As a demonstration of functionality, we analyze crime anomalies in São Paulo, uncovering links to events such as Carnival and religious holidays.
This paper introduces an automated and reproducible pipeline for the generation of synthetic datasets with COCO-format annotations, designed to support object detection and segmentation tasks. The proposed framework combines controlled video acquisition of real-world objects, volumetric reconstruction via the SVRaster algorithm, automated mesh postprocessing in Blender, and synthetic rendering with systematically varied lighting and backgrounds. Annotation generation is fully automated through the custom YOLO FOTO plugin, which orchestrates virtual camera placement and metadata export. The resulting synthetic images are employed to train the YOLOv8n model. The study is conducted under the Design Science Research (DSR) methodology, ensuring methodological rigor and reproducibility. Experimental results demonstrate substantial improvements in model performance when synthetic data are incorporated into training, underscoring their effectiveness in enhancing dataset diversity and robustness. These findings highlight the strategic value of synthetic data in scenarios where manual annotation is costly, time-consuming, or otherwise constrained.
Automated fingerprint identification systems traditionally compare a query against all stored records to identify a match, resulting in a computational burden that is impractical for large-scale fingerprint databases. To address this challenge, indexing techniques are crucial for reducing the search space and accelerating the identification process. This paper introduces a flexible fingerprint indexing framework based on Hierarchical Navigable Small World (HNSW) graphs, a state-of-the-art algorithm for approximate nearest neighbor search. The strength of HNSW relies on its ability to operate beyond metric spaces by supporting arbitrary similarity functions. We demonstrate such versatility through three distinct experimental setups: (1) indexing standard global feature vectors, (2) organizing an extensive collection of local, minutiae-based descriptors aggregated via majority voting, and (3) directly indexing proprietary biometric templates using the native, non-metric matching score of a Commercial-Off-The-Shelf (COTS) system. Our results show that HNSW significantly reduces similarity computations while maintaining high accuracy across all scenarios. Most notably, it minimizes the number of costly COTS matcher evaluations, establishing HNSW as a robust and adaptable solution for automated fingerprint identification systems.
Accurate classification of pulmonary diseases is critical for clinical decision-making, and deep learning models using chest CT scans have become a key tool in this task. Most existing approaches rely on 2D CT slices, which provide limited views and may miss important spatial patterns across the lung volume. To address this, we introduce CT-VR, a novel classification approach that leverages 3D volume-rendered images captured from multiple angles. By incorporating multi-view volume rendering, CT-VR enhances the model ability to detect and differentiate between pulmonary conditions. We evaluate the method using COVID-19 datasets as a primary case study, which include private datasets from partner hospitals and a publicly available benchmark. Results demonstrate that our approach improves lesion identification and delivers performance compared to traditional slice-based models, highlighting its potential as a more effective solution for lung disease classification.
