DRespNeT: A UAV Dataset and YOLOv8-DRN Model for Aerial Instance Segmentation of Building Access Points for Post-Earthquake Search-and-Rescue Missions

Angelos Plastropoulos; Argyrios Zolotas; Aykut Sirma; Gilbert Tang

REVIEW 3 major objections 4 minor 1 cited by

DRespNeT: A UAV Dataset and YOLOv8-DRN Model for Aerial Instance Segmentation of Building Access Points for Post-Earthquake Search-and-Rescue Missions

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A drone-captured dataset with 28 fine-grained classes lets YOLOv8-based models spot accessible doors and windows in post-earthquake rubble at 92.7% mAP50.

desk verdict The new thing here is the dataset, not the model—and the dataset could be genuinely useful if the paper ships the annotation protocol and splits that the abstract omits. read the letter →

arxiv 2508.16016 v2 pith:ZL73LWQA submitted 2025-08-22 cs.CV

Aykut Sirma , Angelos Plastropoulos , Gilbert Tang , Argyrios Zolotas This is my paper · ORCID

classification cs.CV

keywords aerialinstancesegmentationUAVdatasetpost-earthquakesearchandrescuebuildingaccesspointsYOLOv8-segdisasterresponsepolygonannotationsstructuraldamageassessment

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces DRespNeT, a high-resolution (1080p) aerial dataset for instance segmentation of building access points in post-earthquake environments, annotated with polygons across 28 operationally critical classes. The central claim is that this dataset, unlike satellite-derived or coarse semantic alternatives, lets an optimized YOLOv8-based model distinguish accessible from obstructed doors, windows, and gaps, and do so fast enough for live search-and-rescue use. On their benchmark, the YOLOv8-DRN model reports 92.7% mAP50 at 27 FPS on an RTX-4090 GPU. A sympathetic reader would care because knowing which entry points are actually usable is a direct input to rescue routing and robot-assisted search, and existing data does not offer this fine-grained access-point distinction.

What carries the argument

The load-bearing mechanism is the annotation schema itself: DRespNeT's fine-grained polygon-level instance labels on high-definition aerial footage, with classes that separate accessible from obstructed access points. Around that sits the YOLOv8-DRN model, an optimization of the YOLOv8-seg baseline that the paper uses to show the dataset can be consumed in real time. The dataset carries the argument; the model demonstrates the dataset's usability.

What would settle it

Re-annotate a random sample of DRespNeT frames with independent expert labelers and measure agreement on the access-point classes; alternatively, run the trained YOLOv8-DRN on newly collected drone footage from a different post-earthquake site and compare mAP50 with the reported figure. A large drop or low label agreement would show the reported performance is specific to this dataset rather than to post-earthquake scenes generally.

Watch

Extended reading notes

Core claim

On the paper's own terms, the contribution is twofold: DRespNeT supplies polygon-level instance segmentation labels derived from real 1080p drone footage of the 2023 Turkiye earthquake and other impacted regions, covering 28 classes that include compromised buildings, doors, windows, gaps, debris levels, rescue personnel, vehicles, and civilian visibility; and an optimized YOLOv8 variant (YOLOv8-DRN) demonstrates that this dataset supports real-time multi-target detection at 92.7% mAP50. The distinctive claim is that the dataset's fine-grained annotation detail allows automated systems to tell apart accessible and obstructed areas, which is what makes it operationally useful rather than just

Load-bearing premise

The result stands on the assumption that the polygon labels are accurate enough and the footage representative enough that 'accessible' versus 'obstructed' really means what it does in the field; if the labels are noisy or the scenes narrow, the 92.7% mAP50 would not transfer to real rescue operations.

Editorial extensions

If this is right

Search-and-rescue teams could use the dataset to train aerial systems that flag usable doors, windows, and gaps before ground crews arrive.
The 27 FPS inference rate on an RTX-4090 suggests live drone feeds, not just post-hoc analysis, can be processed during a mission.
The 28-class taxonomy gives robotic systems a common vocabulary for distinguishing debris, personnel, vehicles, and civilian visibility in the same frame.
As a shared benchmark, DRespNeT gives later models a concrete target for comparison in post-earthquake structural assessment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the annotations hold up under independent re-labeling, DRespNeT's accessible-versus-obstructed distinction could be transferred to other disaster types, such as floods, fires, or explosions, where the same access-point question arises.
The reported 27 FPS is tied to an RTX-4090; field deployment on smaller drone-mounted computers would likely need model compression, and that is a natural next stress test.
A reader could test whether overall mAP50 hides weaker performance on the safety-critical classes, such as obstructed versus clear doorways, since those classes matter most in rescue decisions.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Worth knowing: DRespNeT is a concrete new artifact—1080p aerial footage from the 2023 Türkiye earthquake and other disaster zones, with polygon-level instance labels for 28 classes including accessible vs. obstructed doorways, windows, gaps, debris levels, and personnel. That is a real gap. Most prior work is satellite imagery or coarse semantic labels, so a fine-grained benchmark for drone-based SAR access-point detection would be useful to an applied niche. The model side is what you would expect: YOLOv8-seg with tweaks, reporting 92.7% mAP50 at 27 FPS on an RTX 4090. Nothing architecturally new, and the abstract does not claim otherwise.

The dataset is the contribution, and I believe it deserves credit. The collection effort and the labeling scheme, if carried out as described, are not trivial. The fact that the same group built both the data and the model is not itself a flaw, but it does mean the number has no external benchmark to anchor it.

The soft spot is exactly where the stress-test note lands: the abstract gives no annotation protocol, no inter-annotator agreement, no dataset statistics, no train/validation/test split, no class-wise results, and no error bars. For a task where 'accessible' vs. 'obstructed' is a judgment call in cluttered disaster footage, a single mAP50 is not a measurement of operational readiness. It shows the model learned something about the labels, not that the labels match what a rescuer would decide. Also, if test frames overlap with training scenes, the number says nothing about generalization to a new earthquake. That concern is central, not minor, but it is about documentation rather than a load-bearing methodological contradiction. mAP50 is a permissive threshold, and one number without variance is over-precise, but none of this makes the dataset idea wrong.

Who is this for? People building drone-based SAR tools and anyone who wants a shared benchmark for fine-grained access-point segmentation. A reader in that community would get value from the dataset itself even if they ignored the performance claim.

My recommendation: take it seriously as a dataset paper and send it to peer review if the full manuscript is available and includes the missing documentation. Desk rejection would be premature because the artifact is genuinely needed. But acceptance should depend on releasing the annotation protocol, inter-annotator agreement, class-wise metrics, and a clear description of the splits so the 92.7% can actually be interpreted.

Referee Report

3 major / 4 minor

Summary. The paper introduces DRespNeT, a high-resolution UAV dataset for aerial instance segmentation of post-earthquake building access points, with 28 operationally critical classes such as structurally compromised buildings, doors, windows, gaps, debris levels, rescue personnel, and civilian visibility. The authors also present a YOLOv8-based model variant, YOLOv8-DRN, and report 92.7% mAP50 at 27 FPS on an RTX-4090 GPU, claiming the system meets real-time requirements for search-and-rescue operations. The contribution is thus a combination of a new fine-grained dataset and a tuned instance-segmentation baseline.

Significance. If the dataset is released with accurate, well-documented polygon annotations and if the benchmark is reproducible, DRespNeT would fill a genuine gap: most existing disaster-response datasets rely on satellite imagery or coarse semantic labels, whereas this work targets fine-grained, instance-level access-point segmentation from low-altitude UAV video. The operational distinction between accessible and obstructed areas is practically valuable for SAR planning and human-robot collaboration. The reported performance is plausible for a tuned YOLO model on a custom dataset, but the abstract alone does not provide enough information to assess its robustness. The strength of the paper lies in the dataset concept and the identification of an under-served application; the quantitative claim is a single benchmark number rather than a derived result.

major comments (3)

[Abstract, performance claim] The sentence 'Our optimized YOLOv8-DRN model achieves 92.7% mAP50 with an inference speed of 27 FPS on an RTX-4090 GPU' is the central quantitative result, but the abstract provides no dataset split, test-set size, class-wise AP, or variance/confidence intervals. mAP50 is a permissive IoU threshold, and a single point estimate without error bars is over-precise for an operational readiness claim. The full paper must specify the evaluation protocol, per-class results, and variability across runs or folds.
[Abstract, dataset documentation] DRespNeT is the paper's primary contribution, yet the abstract gives no annotation protocol, inter-annotator agreement, or dataset statistics (image count, polygon count, class distribution). The distinction between 'accessible' and 'obstructed' access points is an operational judgment in disaster footage; if the labels are subjective or inconsistently applied, the reported mAP only reflects the annotation team's bias. The manuscript must document annotation guidelines, quality control, and independent validation (e.g., agreement with search-and-rescue professionals).
[Abstract, data collection and generalization] Footage from the 2023 Turkiye earthquake and 'other impacted regions' is mentioned, but no train/validation/test split is described. If test frames are drawn from the same scenes or events as training, the 92.7% mAP50 does not demonstrate generalization to new earthquakes. The authors should specify whether the split is by scene, region, and/or event, and report performance on a held-out disaster scenario.

minor comments (4)

[Abstract, terminology] The acronym 'DRN' in YOLOv8-DRN is not defined in the abstract; clarify the architectural modification.
[Abstract, class semantics] The class 'multiple debris levels' is ambiguous; specify whether the levels are ordinal or discrete categories, as this affects annotation semantics.
[Abstract, class naming] The class 'civilian visibility' is unclear: does it denote visible civilians or the visibility condition of civilians? Consider renaming for operational clarity.
[Introduction/Related work (implied)] The claim that existing datasets 'rely heavily on satellite imagery or coarse semantic labeling' should cite and compare against specific alternatives (e.g., xBD, AIDER, Disaster123) in the full text.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity identified; abstract-only review shows standard dataset-and-benchmark structure.

full rationale

The paper is an abstract-only submission. Its central claim is a measured performance result: 'Our optimized YOLOv8-DRN model achieves 92.7% mAP50 with an inference speed of 27 FPS on an RTX-4090 GPU for multi-target detection.' This is an empirical benchmark on the introduced DRespNeT dataset, not a derivation from first principles. No equation or construction is present that would reduce the claimed prediction to an input. The dataset and model are produced by the same group, but that alone is not circularity; the evaluation is against manually annotated ground truth, and there is no evidence that the reported mAP50 is a fitted parameter renamed as a prediction. The abstract lacks annotation protocol, inter-annotator agreement, or dataset statistics, but those are quality/representativeness concerns, not circularity. There is no self-citation chain, no imported uniqueness theorem, and no ansatz smuggled via citation. Consequently, no specific circular step can be identified and the appropriate score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

This paper contributes a dataset and an empirical benchmark rather than a derivation, so the ledger is mostly about the dataset's hidden quality gates. The free parameters are the undisclosed model tuning choices and the evaluation splits, which directly move the reported mAP. The axioms are the assumptions that the labels are correct, the footage represents real SAR conditions, and that a GPU benchmark implies operational readiness. No invented entities are introduced: DRespNeT is an artifact and YOLOv8-DRN is a model variant, not a postulated new phenomenon.

free parameters (2)

YOLOv8-DRN architecture and hyperparameters = not reported in abstract
The tuned variant's architectural changes, loss weights, anchors, and training choices are undisclosed; any tuning on the dataset makes the reported mAP partly a fitted optimum rather than a transfer result.
DRespNeT train/validation/test split and class sampling = not reported in abstract
The evaluation protocol (split strategy, class balancing, image selection from disaster footage) is not disclosed; these choices materially move mAP50 and the generalization claim.

assumptions (3)

domain assumption The 1080p aerial footage is correctly and consistently labeled by expert annotators.
The entire benchmark depends on ground-truth polygon quality; the abstract gives no annotation protocol or inter-annotator agreement.
domain assumption The footage is representative of operational post-earthquake SAR conditions.
Footage from the 2023 Turkiye earthquake and other regions is assumed to cover the distribution of scenes rescue teams will encounter; generalization beyond the collected regions is asserted, not demonstrated.
domain assumption mAP50 on the authors' test set predicts operational utility in the field.
The 'meeting real-time operational requirements' claim converts a GPU benchmark into field readiness, assuming the test set and the RTX-4090 hardware profile match mission conditions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DRespNeT: A UAV Dataset and YOLOv8-DRN Model for Aerial Instance Segmentation of Building Access Points for Post-Earthquake Search-and-Rescue Missions." pith.science (2026). https://pith.science/paper/ZL73LWQA

@misc{pith2026250816016,
  author       = {Pith},
  title        = {Pith review of: DRespNeT: A UAV Dataset and YOLOv8-DRN Model for Aerial Instance Segmentation of Building Access Points for Post-Earthquake Search-and-Rescue Missions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZL73LWQA}},
  note         = {Machine review of arXiv:2508.16016}
}

read the original abstract

Recent advancements in computer vision and deep learning have enhanced disaster-response capabilities, particularly in the rapid assessment of earthquake-affected urban environments. Timely identification of accessible entry points and structural obstacles is essential for effective search-and-rescue (SAR) operations. To address this need, we introduce DRespNeT, a high-resolution dataset specifically developed for aerial instance segmentation of post-earthquake structural environments. Unlike existing datasets, which rely heavily on satellite imagery or coarse semantic labeling, DRespNeT provides detailed polygon-level instance segmentation annotations derived from high-definition (1080p) aerial footage captured in disaster zones, including the 2023 Turkiye earthquake and other impacted regions. The dataset comprises 28 operationally critical classes, including structurally compromised buildings, access points such as doors, windows, and gaps, multiple debris levels, rescue personnel, vehicles, and civilian visibility. A distinctive feature of DRespNeT is its fine-grained annotation detail, enabling differentiation between accessible and obstructed areas, thereby improving operational planning and response efficiency. Performance evaluations using YOLO-based instance segmentation models, specifically YOLOv8-seg, demonstrate significant gains in real-time situational awareness and decision-making. Our optimized YOLOv8-DRN model achieves 92.7% mAP50 with an inference speed of 27 FPS on an RTX-4090 GPU for multi-target detection, meeting real-time operational requirements. The dataset and models support SAR teams and robotic systems, providing a foundation for enhancing human-robot collaboration, streamlining emergency response, and improving survivor outcomes.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Assessing the Benefits of Combining Advanced Deep Learning Techniques for Post-Disaster Building Damage Assessment from UAV Imagery
cs.CV 2026-08 conditional novelty 5.0 of 10

A hybrid CV+LVLM pipeline improves post-disaster building damage counting over single models in some configurations, but fails in others and shows low absolute accuracy.

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.