3DAeroRelief: UAV 3D Disaster Benchmark
- 3DAeroRelief is a 3D benchmark dataset for post-disaster assessment, capturing dense point clouds from low-cost UAV RGB imagery over hurricane-impacted areas.
- It employs a two-stage photogrammetric reconstruction using SfM and MVS, complemented by a multi-view 2D-to-3D semantic annotation pipeline for accuracy.
- The dataset facilitates point-wise semantic segmentation of damage, supporting improved situational awareness and strategic disaster recovery planning.
3DAeroRelief is a 3D benchmark UAV dataset for post-disaster assessment centered on dense point clouds reconstructed from low-cost aerial RGB imagery over hurricane-damaged regions. It was introduced as the first 3D benchmark dataset specifically designed for post-disaster assessment, with the explicit aim of enabling 3D semantic understanding of disaster-affected environments, especially structural damage assessment after hurricanes. The dataset consists of dense 3D point clouds reconstructed via Structure-from-Motion (SfM) and Multi-View Stereo (MVS), with semantic labels obtained by manual 2D annotation, projection into 3D space, and subsequent manual refinement. Its benchmark task is point-wise 3D semantic segmentation under a cross-area generalization protocol (Le et al., 14 Sep 2025).
1. Definition, scope, and dataset composition
3DAeroRelief was introduced to fill a gap between two established but insufficient lines of work. On one side, prior disaster-assessment resources such as FloodNet and RescueNet are based on 2D aerial imagery and therefore lack depth information, explicit 3D spatial context, robustness to occlusions, and fine-grained structural geometry. On the other side, mainstream 3D semantic segmentation benchmarks such as S3DIS, ScanNet, ScanNet++, SemanticKITTI, Waymo, nuScenes, and KITTI-360 focus on indoor scenes or urban driving/street scenes rather than real-world disaster-affected outdoor environments. 3DAeroRelief positions itself at that intersection: large-scale outdoor 3D data, post-disaster structural content, UAV RGB capture, dense photogrammetric reconstruction, and 3D point-wise semantic labels (Le et al., 14 Sep 2025).
The dataset is built from hurricane-damaged regions in Florida affected by Hurricane Ian in 2022. It targets post-disaster assessment tasks including structural damage identification, situational awareness, support for emergency responders, and resource allocation and recovery planning. The acquisition setting is operational rather than purely retrospective: the data were collected during the disaster response phase by the Center for Robot-Assisted Search and Rescue on behalf of the Florida State Emergency Response Team. This operational provenance is central to the dataset’s intended role as a benchmark for robust 3D vision systems in emergency-response settings.
| Property | Value |
|---|---|
| Number of point clouds | 64 |
| Number of areas | 8 |
| Average points per scan | 0.520 million |
| Semantic classes | 5 |
| Disaster event | Hurricane Ian, Florida, 2022 |
| Capture platform | Small UAVs |
| Reconstruction method | SfM + MVS |
The five semantic categories are Building-no-damage, Building-damage, Road, Tree, and Background. The paper defines them precisely as follows: “Buildings with minimal or no visible structural damage” for Building-no-damage, “Buildings exhibiting structural damages” for Building-damage, Road, Tree, and “Remaining unclassified objects and terrain” for Background. The scenes combine damaged buildings, undamaged buildings, roads, trees, and background or terrain, with a stated emphasis on post-disaster built structures.
The dataset is also comparatively dense. Table 1 in the paper reports an average of 0.520 million points per scan, compared with 0.0987 million for SemanticKITTI and 0.041 million for nuScenes. This supports the paper’s claim that 3DAeroRelief captures finer structural detail than street-scene LiDAR datasets.
2. Acquisition, photogrammetric reconstruction, and metric normalization
The raw data consist of high-resolution aerial footage collected by small UAVs over affected Florida regions. Frames were extracted at pixels. The footage was reviewed to remove irrelevant or low-quality content, and the cleaned videos were segmented into individual frames for reconstruction. The paper does not specify UAV model, camera model, focal length, flight altitude, overlap percentages, GPS/IMU specifications, or the total number of extracted or annotated frames; it states only that the platform consists of small UAVs and RGB aerial footage at the stated frame resolution (Le et al., 14 Sep 2025).
The reconstruction pipeline is explicitly two-stage photogrammetry. SfM is first used to recover sparse 3D scene geometry together with camera intrinsics and camera extrinsics. The paper describes the SfM stage in standard incremental terms: keypoints are matched across image pairs to form 2D–2D correspondences; incorrect matches are removed using Homography and Epipolar geometry with RANSAC; reconstruction starts from a well-matched image pair; additional views are added incrementally using Perspective-n-Point; new 3D points are obtained by triangulation; and the model is refined with bundle adjustment minimizing reprojection error. No explicit bundle-adjustment formula is given in the paper.
After sparse reconstruction, MVS densifies the model. The MVS stage includes dense depth estimation using photometric consistency across overlapping views, search-space optimization through plane sweeping and multi-scale filtering, and fusion of depth maps into a final dense point cloud using confidence weighting and consistency checks. The final dataset representation is therefore not LiDAR-based but photogrammetric, reconstructed from UAV RGB imagery via SfM and MVS.
A notable engineering step is the conversion from arbitrary SfM scale to approximate metric scale. The authors manually rescale each reconstruction using the average height of a single-story house, with
Two 3D points corresponding to the house’s vertical extent are selected, the reconstructed height is measured, and a global scale factor is applied: The paper states that this rescaling is necessary because SfM reconstructions have arbitrary scale, while 3D deep networks expect approximately real-world geometry.
The raw reconstructions are further cleaned and cropped to regions of interest. The stated motivations are the removal of large irrelevant areas, reduction of noise, and mitigation of class imbalance caused by an overabundance of undamaged structures. The extracted point-cloud segments were selected to preserve a more representative mix of damaged and undamaged buildings, which also reduces annotation effort.
3. Annotation transfer, semantic representation, and file organization
The annotation strategy is a three-step pipeline: manual 2D image annotation, projection of 2D labels into 3D, and manual 3D refinement. This design is methodologically central because direct manual point-level labeling of dense 3D photogrammetric reconstructions is expensive, especially for disaster scenes with irregular geometry and partial collapse (Le et al., 14 Sep 2025).
In the 2D stage, the authors annotate every 10th frame. The semantic scheme is described as “largely consistent” with RescueNet. The paper does not provide the total number of annotated frames, only the sampling rate. In the projection stage, each 3D point is reprojected into all annotated images using the SfM-estimated camera parameters. The projection relation is given explicitly. For a 3D point
and camera matrix
the corresponding image point is
with . This is the geometric core of the annotation-transfer mechanism.
For each 3D point, valid projections inside image boundaries are collected from all annotated frames, and the final semantic label is assigned by majority voting over the projected 2D labels. The multi-view voting scheme is intended to improve robustness against occlusions, annotation noise, and viewpoint ambiguities. After this transfer step, the resulting 3D labels are manually refined in CloudCompare to correct inconsistencies and improve point-level precision.
The final dataset consists of dense 3D point clouds with per-point semantic labels. The organizational structure is stated explicitly: directories are named Area_<n>, point clouds are named pp<i>.ply, segmentation files are named segmentpp<i>.ply, and labels.txt stores class names and RGB colors. The visualizations use the following color code: Cyan for Building-no-damage, Red for Building-damage, Yellow for Road, Green for Tree, and Black for Background.
This annotation pipeline has two notable implications. First, it ties 3D semantic labels to image-space evidence and camera geometry rather than to purely manual point-cloud labeling. Second, because labels are transferred through multi-view projection and then corrected in 3D, the benchmark reflects both photogrammetric reconstruction quality and semantic annotation quality. A plausible implication is that reconstruction artifacts and annotation uncertainty are structurally entangled in the benchmark rather than separable nuisance factors.
4. Benchmark protocol and evaluation formalism
The benchmark task defined for 3DAeroRelief is 3D semantic segmentation, formulated as point-wise semantic classification over the five classes above. The split is a geographically separated cross-area split designed to test generalization to unseen disaster regions rather than memorization of a single local geometry or texture distribution (Le et al., 14 Sep 2025).
The training set contains 56 point clouds from Areas 1, 3, 4, 5, 6, 7, and 8. The test set contains 8 point clouds from Area 2. The paper does not report a separate validation split. This cross-area setup is stricter than random scene partitioning because the held-out area is geographically distinct.
Evaluation uses Mean Class Accuracy (mAcc) and Mean Intersection over Union (mIoU). The paper gives
and
Equivalent class-wise IoU is written as
0
The evaluated state-of-the-art models are PTv1, PTv2, PTv3, FPT, and OA-CNNs. The paper briefly characterizes them as follows: PTv1 uses self-attention for unordered point sets; PTv2 introduces grouped vector attention, improved positional encoding, and partition-based pooling; PTv3 uses point serialization with grids and space-filling curves and includes the xCPE sparse convolution module; FPT uses centroid-aware voxelization and devoxelization with learnable centroid-to-point positional encoding; and OA-CNNs are omni-adaptive sparse CNNs with adaptive receptive fields and adaptive relation mapping.
A common misconception would be to read 3DAeroRelief primarily as a reconstruction benchmark. The paper’s emphasis is different. Reconstruction is a prerequisite to the dataset, but the benchmark proper evaluates semantic segmentation performance on the reconstructed point clouds. The principal scientific question is therefore not whether a method can build the point cloud, but whether it can semantically parse disaster-specific 3D structure and distinguish damaged from undamaged built forms.
5. Empirical results, class-wise behavior, and domain difficulty
The reported benchmark results show that 3DAeroRelief is substantially harder than a generic outdoor segmentation benchmark. Even recent 3D models exhibit strong class asymmetries and, in some cases, catastrophic failure on specific semantic categories (Le et al., 14 Sep 2025).
| Method | Overall mIoU | Overall mAcc |
|---|---|---|
| PTv1 | 0.3084 | 0.4606 |
| PTv2 | 0.2559 | 0.3941 |
| PTv3 | 0.4584 | 0.5508 |
| FPT | 0.11 | 0.22 |
| OA-CNNs | 0.2604 | 0.4396 |
PTv3 is the strongest overall model, achieving mIoU = 0.4584 and mAcc = 0.5508. The paper attributes this to its serialized representation of point clouds, efficient handling of irregular structures, and ability to combine structural layout with local damage cues. At the class level, PTv3 also achieves the best mIoU for Building-Damage (0.7614), Building-no-Damage (0.3108), Tree (0.4001), and Background (0.7730), while PTv2 is best on Road with 0.4864 mIoU. For mAcc, OA-CNNs is best on Building-Damage (0.9276) and Tree (0.7248), PTv2 is best on Building-no-Damage (0.4826) and Road (0.5152), and PTv3 is best on Background (0.8734).
Several failure modes are especially informative. PTv2 collapses on the Building-Damage class with mIoU = 0.0020 and mAcc = 0.0020. OA-CNNs fails on Road with mIoU = 0.00 and mAcc = 0.00. FPT performs worst overall, with mIoU = 0.11, mAcc = 0.22, and zero IoU for Road and Tree. These are not isolated degradations; they indicate that standard state-of-the-art 3D segmentation architectures do not transfer cleanly to disaster scenes.
The paper identifies several technical reasons for this difficulty. Damaged buildings do not exhibit clean canonical geometry; instead they may involve partial collapse, roof deformation, broken facades, and irregular shapes. Distinguishing Building-Damage from Building-no-Damage is therefore a fine-grained semantic problem that depends on subtle geometric and appearance differences rather than broad object category membership. Occlusion is intrinsic to the scenes and is already implicit in the need for multi-view voting during annotation. Photogrammetric reconstruction introduces outliers, mismatches, noise, and incomplete surfaces. Raw reconstructions also exhibit class imbalance, especially an overabundance of undamaged buildings, which is why the dataset construction isolates representative regions of interest.
This suggests that 3DAeroRelief functions less as a routine transfer benchmark and more as a stress test for 3D semantic robustness under disaster-specific geometric disruption. The spread between damage-class mIoU and mAcc also indicates that point-wise recognition and spatially coherent overlap are not equivalent objectives in this domain. The paper explicitly notes, for example, that OA-CNNs attains the best Building-Damage mAcc while PTv3 attains the best Building-Damage mIoU, which suggests better point coverage for the former and better region overlap for the latter.
6. Relation to adjacent aerial 3D and relief-reconstruction research
3DAeroRelief sits within a broader but heterogeneous literature on aerial 3D perception. Its contribution is not a new reconstruction algorithm, terrain super-resolution method, or localization pipeline; it is a post-disaster UAV benchmark for dense 3D semantic segmentation. That distinction becomes clearer when it is set against nearby problem formulations.
In dense aerial reconstruction, “Aerial Multi-View Stereo via Adaptive Depth Range Inference and Normal Cues” develops ADR-MVS, which improves aerial multi-view depth estimation by using adaptive depth-range prediction and normal-guided processing for dense 3D urban reconstruction (Liu et al., 6 Jun 2025). In terrain enhancement, “AFN: Attentional Feedback Network based 3D Terrain Super-Resolution” formulates high-resolution DEM recovery as a cross-modal super-resolution problem using low-resolution DEM plus co-registered aerial imagery (Kubade et al., 2020). In mission design, “Terrain-Aware Adaptation for Two-Dimensional UAV Path Planners” shows that 2D multi-UAV planners can be post-processed to improve 3D reconstructions of relief and vertical structures through terrain-aware altitude and camera-orientation adjustment (Karakontis et al., 23 Jul 2025). In metric grounding of monocular geometry, “TanDepth: Leveraging Global DEMs for Metric Monocular Depth Estimation in UAVs” uses projected global DEM anchors and ground filtering to recover metric depth from relative monocular predictions (Florea et al., 2024).
A different neighboring strand concerns large-scale cross-platform 3D modeling rather than disaster-scene semantics. “Unconstrained Large-scale 3D Reconstruction and Rendering across Altitudes” introduces a benchmark for camera calibration and novel view synthesis from sparse mixed-altitude imagery, including ground-level, security-level, and airborne cameras, and shows that current baselines fail sharply in the varying-altitudes and reconstructed-area regimes (Joshi et al., 29 Apr 2025). “Aerial-Ground Image Feature Matching via 3D Gaussian Splatting-based Intermediate View Rendering” addresses reliable aerial-ground correspondences by rendering intermediate views from aerial-only 3D Gaussian Splatting, thereby enabling integrated aerial-ground SfM and rendering (Yu et al., 24 Sep 2025). “Altitude-Aware Visual Place Recognition in Top-Down View” addresses a narrower but operationally relevant problem—vision-only altitude normalization for top-down aerial place recognition under large altitude changes—rather than scene reconstruction or semantic parsing (Shao et al., 27 Feb 2026).
Finally, there is a terrain-relief tradition that is adjacent but methodologically distinct from 3DAeroRelief’s disaster benchmark role. “Robust positioning of drones for land use monitoring in strong terrain relief using vision-based navigation” uses a Digital Terrain Map directly for drone localization in mountainous or urban-canyon settings (Kupervasser et al., 2018), while “Ground Profile Recovery from Aerial 3D LiDAR-based Maps” recovers terrestrial relief from UAV LiDAR point clouds using Cloth Simulation Filtering to separate ground from forest returns (Sabirova et al., 2019). These works are directly about relief exploitation or recovery, whereas 3DAeroRelief is about semantic understanding of post-disaster built environments reconstructed from UAV RGB imagery.
The resulting picture is that 3DAeroRelief occupies a distinct position in aerial 3D research. It combines low-cost UAV capture, dense photogrammetric point clouds, post-disaster structural damage classes, and a cross-area 3D semantic segmentation benchmark. It does not replace aerial MVS, terrain modeling, localization, or planning methods; instead, it provides a domain-specific evaluation target for them when the downstream objective is 3D understanding of disaster-affected environments. A plausible implication is that progress on 3DAeroRelief will likely require combining advances from several adjacent areas—better aerial reconstruction, stronger geometric priors, more robust cross-area generalization, and damage-sensitive 3D representations—rather than relying on direct transfer from indoor or autonomous-driving benchmarks alone.