Conference Agenda

This study examines PIX4D’s implementation of Gaussian Splatting for reconstructing complex geometries, with a focus on underwater photogrammetry for coral reef mapping. Unlike standard Gaussian Splatting pipelines that emphasize photorealistic rendering, our approach prioritizes high-precision geometric reconstruction, especially for thin structures and heavily occluded regions. We compare the method against conventional multi-view stereo techniques using both real underwater imagery collected in Moorea (French Polynesia) and synthetic datasets generated with the POSER underwater simulation framework.

Stand-level merchantable tree stem volume estimation in temperate forests is critical for data-driven forest management decision-making. Mobile laser scanning (MLS) has greatly improved data-collection efficiency for forest biometrics; however, automated analysis of massive, structurally complex MLS point clouds remains limited. This study presents an automated framework to estimate stand-level merchantable stem volume from backpack mobile Light Detection and Ranging (LiDAR) data. The framework comprises three stages: (1) point cloud reconstruction using the Integrated-Scan Simultaneous Trajectory Enhancement and Mapping (IS²-TEAM) method; (2) individual tree segmentation via a multistage geometric pipeline; and (3) merchantable stem volume estimation based on skeletonization-derived stem modeling. The proposed approach is evaluated on a forest-scale dataset collected in temperate natural forests in the United States. Results demonstrate operational feasibility at scale, with practical processing times and robust geometric consistency. Validation against destructively measured reference volumes shows that the proposed approach outperforms baseline quantitative structure modeling (QSM) methods, achieving a coefficient of determination (R²) of 0.97, a bias of −0.06 m³, and a root mean square error (RMSE) of 0.21 m³. The proposed framework enables reliable, automated estimation of merchantable stem volume from MLS data and supports deployment from individual-tree to forest scales with minimal manual intervention.

TRACE is a training-free framework for instance-level open-vocabulary inventory generation in 3D forensic evidence reconstruction.

Starting from multiview RGB imagery, prompt-based 2D object masks are extracted using SAM3 and associated across views via geometry-aware and appearance-aware multiview instancing. Based on COLMAP geometry and DINOv2/v3 descriptors, the proposed framework establishes globally consistent same-class object identities across the scene. The resulting global instances are then encoded with SigLIP2 to obtain language-aligned instance descriptors and subsequently lifted into a 3D Gaussian Splat representation by assigning instance-level semantics to geometrically supported Gaussian subsets. This yields an enriched 3D scene representation that jointly preserves spatial structure, object-level identity, and language-accessible semantics, thereby enabling instance-aware open-vocabulary querying in 3D.

The purpose of this study is to automatically classify water and land areas with LiDAR point clouds. After determining the average water level, the water and land surfaces were classified. Previous studies have focused on supervised classification based on land sampling or deep learning techniques using photographs. However, these classification techniques are expensive and require long calculation times. In this study, a method is proposed for the automatic classification of water and land areas without land surveys using the coordinate and reflection values of LiDAR point clouds. The bounding box method was used to detect water surface levels. The correlations between the min-box level, mean box height, and mean box reflection values of the LiDAR point data were used to determine the water surface level. The results show that the method is suitable for the fast classification of water surfaces from LiDAR point clouds. Thus, shoreline changes in large areas can be detected automatically without the need for land surveying. The proposed bounding box classification method can be applied independently of LiDAR point cloud density. The extended version of this method can also be used to detect vehicles and objects on a water surface.

Underground environments such as sewer networks are critical infrastructure whose condition directly affects public health, environmental protection, and maintenance costs. Conventional inspection workflows largely rely on monocular CCTV systems and manual video review, providing limited 3D understanding and often missing subtle or spatially complex defects. At the same time, sewer environments are characterised by challenging imaging conditions, including low illumination, specular surfaces, water films and occlusions, which further complicate reliable assessment.

In this extended abstract, we present a real-time inspection concept that combines (i) stereo camera-based SLAM for geometric mapping and pose estimation, (ii) Vision Transformer (ViT) based anomaly detection trained on the public SewerML dataset, and (iii) lightweight Gaussian Splatting modules that create local high-resolution 3D reconstructions only in the vicinity of detected defects. The system is targeted at embedded hardware, specifically an NVIDIA Jetson Nano, and is designed for deployment and evaluation in real sewer environments. The overall goal is to provide inspectors and asset managers with spatially anchored 3D visualisations of anomalies that can be integrated into digital-twin workflows for decision support and long-term monitoring.

This paper proposes a robust cross-modal matching framework for tunnel inspection, specifically designed to address the unique challenges posed by low-texture environments often encountered in tunnel linings. Traditional image-based matching techniques struggle in these environments due to the lack of distinctive surface features and limited texture variation. To overcome these challenges, the proposed method leverages the global prior knowledge of tunnel geometry. By jointly projecting LiDAR point clouds and multi-camera images onto a shared parameterized cylindrical surface, the method constructs a unified geometric space that facilitates accurate 3D–2D correspondences. This dual-projection strategy significantly improves the alignment of structural features such as segment joints, line grooves, and equipment brackets, which are critical for defect detection in tunnel inspection. The enhanced matching ability allows for more reliable multi-sensor data fusion, thereby supporting the automated analysis of tunnel defects. This framework lays a solid foundation for intelligent tunnel inspection systems, offering a powerful solution for real-time monitoring and analysis of tunnel infrastructure.

The AQpictures project, conducted under the ISPRS Scientific Initiatives 2025, addresses the emerging field of image-based estimation of fine particulate matter (PM2.5) concentrations in urban areas. PM2.5 represents a major public health concern, yet existing ground-based monitoring networks offer limited spatial coverage and satellite-derived products struggle to capture surface-level variability. Recent studies have demonstrated that visual attributes in outdoor images, such as sky colour, haze, and visibility, can provide useful indicators of PM2.5 concentrations. Building upon this premise, AQpictures aims to develop an open, reproducible framework for benchmarking and validating image-based air quality estimation methods.

The project first conducts a comprehensive literature review to classify existing approaches into four methodological categories: physics-based, machine learning, deep learning, and hybrid models. Based on this synthesis, a benchmark experiment is implemented for the city of Milan, combining a ten-month dataset of webcam images with co-located PM2.5 ground measurements. The workflow involves image preprocessing, feature extraction, and model evaluation using standard statistical indicators (R², RMSE, MAE). Preliminary tests include physics-based visibility models, feature-based regressors, and convolutional deep learning architectures.

All codes, datasets, and documentation are consolidated in an open-access GitHub repository to ensure transparency, reproducibility, and adaptability of methods across different environmental contexts. Early results confirm the feasibility of PM2.5 estimation from RGB imagery, though further investigations on multi-city datasets are planned to evaluate model transferability and robustness under varying urban and climatic conditions.

This paper presents MAMAI (Maternal Assistance and Monitoring through Artificial Intelligence), a standards-based framework designed to enable interoperable postpartum well-being monitoring using multi-vendor wearable devices. The proposed system addresses a key limitation in digital maternal health: the fragmentation of wearable ecosystems and the lack of integration with clinical infrastructures. MAMAI introduces a federated, edge–cloud architecture that allows wearable data to be processed locally while transmitting only summarized to the cloud.

A core contribution of this work is the integration of two complementary interoperability standards: the OGC SensorThings API for structuring IoT-based sensor observations, and HL7 FHIR for representing well-being indicators in clinically compatible formats. Through this dual-standard approach, heterogeneous wearable data—such as sleep patterns, physical activity, and heart-rate variability—are harmonized into standardized, platform-independent representations.

The framework further introduces a composite well-being score derived from normalized physiological indicators, enabling continuous and interpretable assessment of maternal health. A prototype implementation demonstrates the feasibility of the architecture, supporting end-to-end data ingestion, transformation, interoperability mapping, and visualization. Experimental results show efficient system performance with low end-to-end latency.

Overall, MAMAI provides a scalable and interoperable solution for integrating consumer wearable data into healthcare ecosystems, offering a foundation for next-generation maternal digital health systems and continuous postpartum monitoring.

Achieving the United Nations Sustainable Development Goal (SDG) 11.7.1—“providing universal access to safe, inclusive, accessible, and green public spaces by 2030”—underscores the critical role of urban green space in advancing global sustainability.Although extensive research has examined urban greenery from a traditional planar perspective, green spaces inherently possess vertical structure. Currently, systematic quantitative assessments of urban vertical greenery, residents’ actual exposure to vertical green space, and the associated inequalities remain limited. To address these gaps, this study integrates global population data with vegetation height information to construct an exposure-based analytical framework.We quantify spatial patterns of vertical greenery, residents’ green exposure, and exposure inequality across global urban areas, and further examine the drivers of inequality. Our findings reveal pronounced spatial disparities in urban greenery worldwide. On average, cities in the Global North exhibit approximately three times greater vertical greenery and nearly four times higher green exposure than cities in the Global South. African urban areas possess only one-sixth of the average vertical greenery and one-seventh of the exposure level observed in North America, while displaying roughly twice the inequality in green exposure, indicating much more uneven access to green resources. We also find that cities with higher average vertical greenery tend to experience lower exposure inequality, suggesting that increasing overall greenery can help promote more equitable access. These results provide new theoretical insights and policy-relevant evidence for advancing sustainable and equitable urban green development, supporting global progress toward sustainable development goals.

Walkability assessments for elderly pedestrians are often based on static representations of the built environment, overlooking temporal variations that influence walking conditions throughout the day. This study develops a network-based dynamic walkability framework that integrates static infrastructural characteristics with time-dependent environmental factors to capture spatiotemporal variability in pedestrian suitability. The approach combines sidewalk and arcade-based pedestrian networks with dynamic variables, including traffic, air quality index (AQI), temperature, humidity, shade, and lighting, evaluated at two time periods (12:00 and 17:00) across weekdays and weekends in three urban contexts in Tainan, Taiwan: a hospital area, a university campus, and a residential neighborhood. Results indicate clear spatial differences, with hospital and campus areas showing higher baseline walkability than residential areas. Dynamic analysis reveals temporal variation, with improvements ranging from approximately 3–8% in institutional environments to over 10% in residential areas. Segment-level results further show that temporal factors can alter pedestrian suitability, particularly in areas with limited infrastructure. Route-based validation demonstrates that the model generates alternative paths that prioritize safety and environmental comfort over the shortest distance. Compared to Google Maps routes, the proposed approach achieves higher average walkability, with improvements ranging from approximately 5% to over 15%, particularly in residential areas. These findings highlight the limitations of static and shortest-path approaches and emphasize the importance of incorporating temporal dynamics. The proposed framework supports time-sensitive routing and age-friendly urban planning strategies.

Existing building digital twins treat indoor environments as static geometric containers, ignoring the dynamic coupling between ventilation structure states and indoor environmental quality. Furthermore, managing indoor and outdoor spaces as separate data silos prevents the continuous assessment of occupant exposure across building boundaries. This paper proposes an environment-aware, indoor-outdoor integrated digital twin framework coupling geometric entity states with physical environmental fields for healthy mobility assessment. The framework utilizes a three-layer architecture. First, the Geometric-Semantic Layer provides a seamless LOD4 model with topologically stitched spaces, modeling ventilation facilities as first-class entities with mutable state attributes (Full Closed, Half Open, Full Open). Second, the Physical Field Layer maps mobile sensing data (PM2.5, CO2) onto semantic entities using a semantic-constrained method, treating walls and closed windows as aggregation barriers. Finally, the Behavioral Response Layer combines entity-level pollution values with pedestrian counts to compute a cumulative Crowd Exposure Index (CEI). Implemented on a Cesium platform, the framework was validated through a week-long university building experiment. Results show indoor PM2.5 in a fully enclosed study room averaged 61.2 μg/m³—1.6 times the outdoor level and 4.1 times the WHO guideline. This resulted in a CEI 12 times higher than in outdoor transit areas. Semantic correlation confirms the "Full Closed" window state primarily drives pollutant accumulation. This validates the framework's core geometry-physics coupling, demonstrating its potential to guide intelligent ventilation interventions and healthy building management.

Rapid industrialization and unplanned urbanization have increased air pollution levels across Indian cities, posing serious environmental and health challenges. This research presents a geospatial assessment of air pollutant behaviour across Nashik city by integrating multi-source remote sensing datasets and real observation datasets from Sentinel-5P, NASA POWER, and CPCB ground observations within a GIS-based analytical framework. Using ward-level mapping and spatial overlays, the study examines the distribution of key pollutants—PM2.5, PM10, NO2, SO2, and CO—and their relationship with environmental and anthropogenic parameters, including land use, road networks, wind direction, temperature, and vegetation density. The results consistently reveal high concentrations of PM2.5, ranging from a minimum of 52.4 µg/m³ to a maximum of 73 µg/m³, and PM10, a minimum of 87.3 µg/m³ and a maximum of 121.5 µg/m³, particularly along high-traffic corridors and industrial zones, which exceed the WHO standards. Correlations with meteorological and vegetative factors further highlight the influence of urban form and climatic conditions on pollutant dispersion. This integrated approach demonstrates how multi-source remote sensing and GIS tools can be effectively employed to identify emission hotspots, support evidence-based policy formulation, and strengthen urban environmental management strategies for sustainable development.

Air pollution is a critical environmental challenge, with nitrogen dioxide (NO₂) from vehicles and industries posing serious health and atmospheric risks. Traditional monitoring is limited, making satellite-based methods essential for large-scale assessment. Korba, Chhattisgarh is an industrial hub of coal mining and thermal power plants is a major pollution contributor. This study investigates the spatiotemporal dynamics, statistical behavior, and long-term trends of NO₂ concentrations over the Korba region from 2019 to 2024, utilizing Sentinel-5P TROPOMI-derived NO₂ column density and the Normalized Difference Pollution Index (NDPI). Year-wise NDPI patterns revealed a consistent pollution hotspot in the central-southern region, with the annual mean NDPI gradually increasing from 0.175 in 2019 to 0.191 in 2023. The monthly NDPI peaked in December-2024 at 0.525, indicating severe winter pollution. Statistical analysis showed moderate variability and a near-symmetric NDPI distribution with occasional spikes near industrial zones. Trend analysis identified a marginal but steady increase in pollution. Autocorrelation analysis revealed strong short-term persistence (lag-1 = 0.594), while spectral analysis identified a dominant annual frequency (0.083 cycles/month) with a peak power of 0.107, confirming the presence of strong seasonal variation and short-term persistence in NO₂ concentration. These results underscore the cyclic yet escalating nature of NO₂ pollution, with notable winter intensification. The findings emphasize the need for targeted emission control strategies and policy-level interventions to manage regional air quality. Future work should integrate ground-based validation and explore meteorological influences to improve predictive accuracy and guide sustainable environmental management.

Multi-view image-to-point label transfer is an effective strategy for 3D semantic segmentation, but its performance largely depends on how predictions from multiple image observations are fused for each 3D point. Most existing pipelines rely on hard voting or handcrafted weighting rules, which do not explicitly learn the reliability of each view under varying geometric and image-quality conditions. In this paper, we introduce DeepChoice, a lightweight view-weighting module for image-guided 3D semantic segmentation. For each visible observation of a 3D point, DeepChoice exploits a compact set of visibility cues, including incidence angle, range, contrast, sharpness, signal-to-noise ratio, and saturation, to predict normalized per-view weights used to aggregate 2D semantic class probabilities into final 3D point-wise predictions. The method is sensor-agnostic, requires no meshing, and can be integrated as a replacement for standard multi-view fusion rules. Experiments on the full GridNet-HD benchmark show that DeepChoice improves over hard voting by 3.85 mIoU points and over mean-probability fusion by 1.26 points, while reducing the gap with the AnyView oracle upper bound. The largest gains are observed on thin and difficult classes such as conductors, pylons, and insulators. Furthermore, a complementary evaluation on the Images PointClouds Cultural Heritage}dataset shows that the proposed weighting strategy remains beneficial under a very different acquisition context and scene structure, yielding a 1.55 mIoU point improvement over hard voting. These results show that learning how to weight views is a simple yet effective way to strengthen image-guided 3D semantic segmentation pipelines. Code is publicly available at: https://huggingface.co/heig-vd-geo/DeepChoice.

Textured 3D meshes jointly encode geometry, topology, and appearance, yet their irregular structure poses significant challenges for deep-learning-based semantic segmentation.

While a few recent methods operate directly on meshes without imposing geometric constraints, they typically overlook the rich textural information also provided by such meshes. We introduce a texture-aware transformer that learns directly from raw pixels associated with each mesh face, coupled with a new hierarchical learning scheme for multi-scale feature aggregation.

A texture branch summarizes all face-level pixels into a learnable token, which is fused with geometrical descriptors and processed by a stack of Two-Stage Transformer Blocks (TSTB), which allow for both a local and a global information flow.

We evaluate our model on the Semantic Urban Meshes benchmark and a newly curated cultural-heritage dataset comprising textured roof tiles with triangle-level annotations with damage types.

Our method achieves 81.9\% mF1 and 94.3\% OA on SUM, and 49.7\% mF1 and 72.8\% OA on new dataset, substantially outperforming existing approaches.

Automatic pothole detection is important for improving road maintenance and transportation safety. While image-based pothole detection often struggles under poor lighting and weather conditions, point cloud data provides a robust alternative by capturing detailed surface geometry. Machine learning has demonstrated strong performance in point cloud classification. While traditional machine learning is simpler and relies on handcrafted features, deep learning models are more powerful, as they learn complex, high-dimensional patterns directly from the input data. While most existing work relies on deep learning models, which are time-consuming to train and require extensive labelled datasets, potholes can be well described by geometric features, making pothole detection well-suited for feature engineering. This paper compares traditional machine learning and deep learning approaches for pothole classification using point cloud data, to evaluate whether the added complexity and data demands of deep learning models are justified, or if traditional machine learning techniques are sufficient for accurate classification. A dataset with labelled pothole instances is created to train both models. The machine learning approach uses manually engineered geometric features as input to an ensemble classifier, while the deep learning model is trained on sampled data. Experimental results show that the machine learning approach outperformed the deep learning model. These results suggest that for this particular task, where informative domain-specific features can be manually engineered, the machine learning approach offers a more practical and efficient solution for real-world deployment, where labelled data may be limited.

Synthesizing realistic fac¸ade views of individual trees from nadir-view remote sensing data would transform large-scale forest

analysis, yet remains unsolved due to data scarcity and task ambiguity. We present the first conditional diffusion model to generate

structurally plausible fac¸ade views of individual tree crowns from single nadir-view LiDAR rasters, leveraging the FOR-species20K

benchmark dataset. Our approach integrates nadir projections with tree species and height within a U-Net-based denoising diffusion

framework. Experiments demonstrate that nadir imagery alone is insufficient, but conditioning on species and height enables

synthesis of visually realistic, species-specific fac¸ade views. The fully conditioned model achieves substantial gains in perceptual

(LPIPS: 0.184) and structural (SSIM: 0.576) similarity, outperforming nadir-only baselines by more than twofold. Our results

establish that ancillary attributes critically constrain the solution space, enabling diffusion models to infer plausible structures

from ambiguous nadir input. This work demonstrates a scalable path to enriching nadir-based forest inventories with synthesized

structural detail, reducing the need for resource-intensive ground surveys.

Metric monocular depth estimation has become increasingly important and is often used as a redundancy mechanism in autonom

ous driving, where accurate scene understanding is essential for safe decision-making. In this work, we evaluate three recently

proposed models that represent the state-of-the-art (Depth Anything, PackNet-SfM, and UnidDepth) using zero-shot testing on the

DrivingStereo dataset across diverse weather conditions, and benchmark their performance. Our analysis considers not only metric

depth accuracy metrcis but also each model’s ability to generalize under challenging environmental variations. While UniDepth

achieves notable improvements over Depth Anything and PackNet-SfM, our results show that substantial progress is still needed for

robust real-world deployment. To further assess its practical suitability for autonomous driving applications, we conduct a detailed

examination of UniDepth’s strengths, limitations, and failure modes.

We present the first Open-Set Recognition (OSR) workflow for environmental monitoring for Permanent Laser Scanning (PLS) setups, using a Deep Neural Network (DNN) solely trained on simulated data. Such monitoring systems were previously only trained with real-world data and under the closed-set assumption, because they are commonly designed to observe a specific and predefined phenomenon (e.g., beach erosion, rockfall activity, vegetation change, animal behavior). The use of real-world data requires manual labeling, which is tedious given the great amount of point clouds. For this reason, we use Virtual Laser Scanning of Dynamic Scenes (VLS-4D) in a PLS setup to investigate how knowledge from synthetic data can be applied to real-world PLS monitoring systems in open-set settings. We introduce a novel framework that enables Open-Set Recognition (OSR) for animal monitoring (e.g. honey bees) using PLS data. The DNN is fine-tuned exclusively on a simulated LiDAR point cloud time series of flying honey bees, and integrates OSR to handle unknown classes during real-world deployment (e.g., butterflies, leaves, wren, and hare). By leveraging deviations in feature embeddings of the DNN, our method reliably distinguishes the known honey bee class from previously unseen classes, supporting robust monitoring under persistent distribution shifts. This approach reduces the dependence on extensive manual annotation of real-world point clouds, while maintaining reliable classification performance. It also highlights the potential of synthetic training data and OSR for environmental monitoring with PLS systems.