Conference Agenda

Real-time depth reconstruction from ultra-high-resolution UAV imagery is essential for time-critical geospatial tasks such as disaster response, yet remains challenging due to wide-baseline parallax, large image sizes, low-texture or specular surfaces, occlusions, and strict computational constraints. Recent zero-shot diffusion models offer fast per-image dense predictions without task-specific retraining, and require fewer labelled datasets than transformer-based predictors while avoiding the rigid capture geometry requirement of classical multi-view stereo. However, their probabilistic inference prevents reliable metric accuracy and temporal consistency across sequential frames and overlapping tiles. We present ZeD-MAP, a cluster-level framework that converts a test-time diffusion depth model into a metrically consistent, SLAM-like mapping pipeline by integrating incremental cluster-based bundle adjustment (BA). Streamed UAV frames are grouped into overlapping clusters; periodic BA produces metrically consistent poses and sparse 3D tie-points, which are reprojected into selected frames and used as metric guidance for diffusion-based depth estimation. Validation on ground-marker flights captured at approximately 50 m altitude (GSD ≈ 0.85 cm/px, ~2,650 m² ground coverage per frame) with the DLR Modular Aerial Camera System (MACS) shows that our method achieves sub-meter accuracy, with approximately 0.87 m error in the horizontal (XY) plane and 0.12 m in the vertical (Z) direction, while maintaining per-image runtimes between 1.47 and 4.91 seconds. Results are subject to minor noise from manual point-cloud annotation. These findings show that BA-based metric guidance provides consistency comparable to classical photogrammetric methods while significantly accelerating processing, enabling real-time 3D map generation.

Satellite-based 3D reconstruction has gained prominence with the advancement of Earth Observation techniques. Recent work on Earth Observation Gaussian Splatting (EOGS) demonstrated the potential of adapting 3D Gaussian Splatting to satellite imagery, enabling rapid Digital Surface Model (DSM) generation from multiple images using Rational Polynomial Coefficients (RPCs) as camera models. However, EOGS suffers from critical inefficiencies: it randomly initializes a large number of Gaussians in volumetric space and relies on opacity-based pruning, resulting in unstable memory footprints and premature loss of fine details—particularly problematic for low-resolution satellite data.

This work presents an improved Gaussian Splatting framework for satellite imagery that addresses these limitations through two key contributions. First, we introduce bundle-adjusted initialization, which leverages geometrically precise points from the bundle adjustment process as initialization seeds rather than random placement. This approach ensures Gaussians are anchored to accurate geometric positions from the outset, significantly improving convergence stability. Second, we propose densification-included optimization, which strategically adds Gaussians in regions requiring detailed reconstruction while maintaining computational efficiency. This selective densification preserves fine-scale features without the memory overhead of EOGS's initial over-allocation strategy.

Our method achieves faster processing times and maintains more consistent memory usage while producing higher-quality DSMs, particularly in challenging low-resolution scenarios. By combining geometric priors from bundle adjustment with adaptive densification, we enable more practical and efficient satellite-based 3D reconstruction suitable for large-scale Earth observation applications.

This contribution presents a comparative evaluation of classical and deep learning–based keypoint detectors for Structure-from-Motion (SfM) reconstruction in challenging Arctic UAV imagery. Snow-covered environments pose difficulties for standard feature matching due to low texture, repetitive patterns, and specular surfaces. While deep keypoint pipelines have shown strong performance on indoor and urban benchmarks, their effectiveness in winter aerial domains remains largely unexplored.

Using multi-view UAV datasets collected across several Alaskan sites, we benchmark three feature-extraction front-ends within a uniform pycolmap-based SfM pipeline: (i) classical SIFT with nearest-neighbor matching; (ii) SuperPoint, a self-supervised convolutional detector–descriptor; and (iii) DISK, a reinforcement-learning–based feature extractor. A simple hybrid approach combining SuperPoint and DISK matches is also tested. All methods share identical geometric verification and bundle-adjustment settings to ensure consistency.

Results show that SIFT remains highly robust on moderately textured Arctic scenes, registering all images and producing the most complete point clouds. SuperPoint and DISK achieve similar reprojection accuracy but struggle with image registration and keypoint coverage on some sequences. Conversely, on extremely low-texture scenes where SIFT fails almost entirely, both deep methods still enable partial reconstructions. Persistent failure cases for all techniques include dense canopy and homogeneous snow.

The study highlights a domain gap between existing deep keypoint models and Arctic aerial imagery, suggesting that domain-specific training and improved spatial keypoint diversity could substantially enhance deep SfM performance in polar regions.

Accurate 3D reconstruction is a key enabler for construction progress monitoring and digital-twin maintenance. However, in tower-crane imagery, persistent dynamic occluders such as hooks and slings violate the static-scene assumption of conventional Structure-from-Motion (SfM), leading to feature mismatches and degraded reconstruction consistency. In this paper, we present a geometry-guided occlusion-handling pipeline for crane-mounted construction-site SfM. Our approach leverages geometric cues from reprojection errors and depth inconsistencies to identify outlier observations, clusters them into spatially coherent prompts, and uses these to guide a foundation segmentation model (SAM2). The resulting per-frame masks are integrated into mask-constrained SfM optimization, ensuring that only static background contributes to reconstruction. Experiments on three real-world crane-mounted sequences (30m, 45m, and 120m) show consistent reductions in mean reprojection error relative to the unmasked baseline. In the most challenging case, the error decreases from 0.962 to 0.872 pixels (9.4%). Compared with a fixed rectangular masking strategy, the proposed masks yield similar reprojection errors while better preserving valid observations and sparse-point completeness. These results indicate that the proposed framework provides a practical geometry-guided strategy for improving internal reconstruction consistency in crane-mounted construction environments.

Video panoptic segmentation (VPS) unifies panoptic segmentation and object tracking by assigning each pixel a semantic class label, or for thing classes, an instance identifier that is consistent across frames. Addressing this task, we propose a novel online VPS method for processing stereoscopic image sequences, which is based on depth-aware kernel-based panoptic segmentation. Specifically, we introduce a geometrical constraint based on predicted bounding boxes into the segmentation of thing instances to overcome the fundamental limitation of kernel-based panoptic segmentation that only appearance information is considered in this step; this regularly leads to panoptic segmentation results in which distinct instances are erroneously merged into one mask. To link detected instances across frames, we propose to extend the commonly employed appearance-based association with a motion-related constraint based on optical flow; this resolves ambiguities in case of instances of similar appearance and, thus, reduces the number of incorrect associations. We experimentally evaluate our method on the publicly available Cityscapes-VPS dataset and compare our results to those of several related methods from the literature. The results demonstrate that our method improves the panoptic quality for a single frame and enhances the instance association across frames, leading to an overall improvement of 3.5% in Video Panoptic Quality on thing classes compared to the employed baseline.

This study addresses the morphological evolution of the Belvedere Glacier (Monte Rosa, Macugnaga – Italy) over the period 2009–2023, using a photogrammetric methodology based on Pleiades (2017) and Pleiades Neo (2023) Very-High Resolution (VHR) satellite imagery, integrated with historical aerial data from 2009. The main objective was to quantify altimetric and volumetric variations of the glacier, assess the intensity of ice mass loss, and analyze the geomorphological effects of the flood event that occurred on August 27, 2023, which generated a major debris flow.

Raster differencing between Digital Elevation Models (DEMs) revealed a significant lowering of the glacier surface. Between 2009 and 2017, the glacier lost approximately 19.3 × 10⁶ m³ of ice (about 2.4 × 10⁶ m³/year), while in the following period (2017–2023) the loss reached 16.9 × 10⁶ m³, with an increased average annual rate of 2.8 × 10⁶ m³/year. These values confirm an acceleration in the ablation process, consistent with other studies (De Gaetani 2021; Ioli 2023) and with the general retreat trend observed in Alpine glaciers due to climate warming.