Pith. sign in

REVIEW 2 major objections 5 minor 18 references

Unlocking Thermal Aerial Imaging: Synthetic Enhancement of UAV Datasets

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a procedural pipeline can render synthetic thermal objects onto real aerial backgrounds, adding new detection classes with drone mAP@0.5 of 0.995 and deer mAP@0.5 of 0.930.

desk verdict Useful dataset pipeline, but the headline mAP numbers are synthetic-to-synthetic consistency, not real-world transfer. read the letter →

arxiv 2507.06797 v1 pith:JGYC5EOM submitted 2025-07-09 cs.CV

Antonella Barisic Kulas , Andreja Jurasovic , Stjepan Bogdan This is my paper · ORCID
classification cs.CV
keywords syntheticthermalimagingUAVdatasetsproceduraldataaugmentationaerialobjectdetectionshaderrenderingorientedboundingboxessegmentationmasksinfraredsimulation
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

This paper claims that a procedural synthetic-image pipeline can add entirely new object classes to existing thermal aerial datasets without collecting new real thermal footage. The authors build 3D models with custom thermal-shader materials, render them from viewpoints matched to the metadata of real thermal backgrounds, overlay the renders, and auto-generate axis-aligned boxes, oriented boxes, and masks. On the resulting datasets, HIT-UAV-drone and MONET-deer, a YOLOv8 detector reaches mAP@0.5 of 0.995 for drones and 0.930 for deer while existing classes keep roughly their original performance. The intended payoff is a cheap way to expand thermal UAV perception to arbitrary classes, useful for search-and-rescue, wildlife monitoring, and urban air mobility.

What carries the argument

The load-bearing object is the procedural rendering pipeline in a 3D editor: thermal backgrounds with per-image metadata define the virtual camera; a custom InfraredEffect shader node, tuned for each object, assigns heat-like emission to the model's surfaces, with hot motors and batteries on the drone and a more uniform body temperature on the deer; the Cycles engine path-traces the scene; and the rendered layer is alpha-overlaid onto the real background. The same projection used for rendering writes axis-aligned bounding boxes, oriented bounding boxes, and masks automatically. The second core component is metadata alignment: the virtual camera reproduces the recorded angle of each background, which the ablation shows is what makes the synthetic objects look like part of the aerial thermal scene.

What would settle it

Collect real aerial thermal images containing drones in urban settings and deer in rural fields, annotate them, and run the paper's HIT-UAV-drone and MONET-deer trained detectors on those images without fine-tuning; if mAP@0.5 is substantially below the reported 0.995 and 0.930, the synthetic-transfer claim fails. A cheaper check is to measure the thermal contrast statistics of the synthetic objects against real thermal imagery of the same classes: if the rendered object temperatures lie outside the real distribution, transfer is expected to degrade.

Watch

Extended reading notes

Core claim

The central claim is that a single procedural pipeline, built on real thermal backgrounds plus Blender-rendered synthetic objects with manually designed infrared shaders, can extend an existing thermal aerial dataset with new classes, oriented boxes, and masks at no extra labeling cost, and that detectors trained on the augmented data perform strongly. The paper reports drone mAP@0.5 of 0.995 on HIT-UAV-drone and deer mAP@0.5 of 0.930 on MONET-deer, while original classes remain near their previous metrics. An ablation shows that aligning the virtual camera pitch with real background metadata outperforms a fixed frontal view, 0.995 versus 0.464 mAP@0.5, and that a visible-light-trained detector transfers poorly to thermal test images, 0.429 versus 0.995.

Load-bearing premise

The approach assumes that manually authored thermal shader nodes make synthetic drones and deer look enough like real thermal objects that a detector trained on synthetic-overlaid images will recognize real aerial thermal footage of those classes, and the reported experiments do not test that transfer because both training and test images use the same 3D models and rendering pipeline.

Editorial extensions

If this is right

  • New thermal object classes can be added to an existing dataset from 3D models and rendered overlays only, without new field captures or manual annotation.
  • Oriented bounding boxes and segmentation masks for the new class are produced automatically from the 3D scene, enabling training of oriented-detection and segmentation models at no labeling cost.
  • Thermal-trained detectors will outperform visible-light-trained detectors on thermal test data for the same class, because the modality gap dominates appearance.
  • Matching synthetic camera pitch to the real background's recorded angle is important; fixed frontal rendering degrades new-class mAP@0.5 from 0.995 to 0.464 in the paper's setup.
  • The dataset can be doubled in size by rendering multiple scenes per background, so the method also expands data quantity, not just class coverage.

Reading between the lines

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

  • The reported mAP values measure how well a detector recognizes synthetic objects rendered by the same pipeline, since both train and test splits for drone and deer come from the same 3D models and renderer; the real test, which the paper does not perform, is whether these detectors transfer to real aerial thermal footage of drones and deer.
  • Because synthetic annotations are generated by projection, label noise on the new class is effectively zero, so real-world performance would likely be lower than the in-pipeline metrics suggest.
  • The pipeline's main bottleneck is the manual authoring of a credible thermal shader for each new object; if thermal appearance can be approximated procedurally from CAD models, the approach extends to classes like power lines, towers, or other wildlife species named in the paper's future work.
  • A natural extension is closed-loop dataset generation: render only objects that confuse the current detector or that maximize its uncertainty, making the overlay process adaptive to the trained model's weaknesses.
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.

Referee Report

2 major / 5 minor

Summary. The paper proposes a Blender-based procedural pipeline that overlays synthetic 3D objects (a drone and a deer) onto real thermal aerial backgrounds from the HIT-UAV and MONET datasets, with the virtual camera aligned to the recorded metadata of each background. The pipeline automatically produces axis-aligned bounding boxes, oriented bounding boxes, and segmentation masks, and the authors release two enhanced datasets (HIT-UAV-drone and MONET-deer). They train YOLOv8s on these datasets and report high mAP for the new classes (drone 0.995, deer 0.930), preserved performance on existing classes, and ablations comparing visible-light versus thermal training and fixed, random, versus metadata-aligned camera angles. The central claim is that this pipeline validates a low-cost route to expanding thermal aerial datasets to new object classes and new applications.

Significance. If the pipeline were shown to transfer to real thermal data, it would be practically significant: it would allow researchers to add new classes, oriented boxes, and masks to existing thermal aerial datasets without new data collection. The manuscript is clearly written, the method is described in sufficient detail to be reimplemented, the public release of the code and datasets is a strength, and the internal check that adding a new class does not degrade existing classes (Tables III and VI) is useful. The angle-alignment ablation (Table VII) is also a reasonable internal consistency test. However, the evidence for real-world transfer is absent: all new-class test images are generated by the same rendering pipeline as the training images, so the headline mAP values do not measure out-of-distribution generalization. The practical significance of the work is therefore conditional on external validation.

major comments (2)
  1. [Section IV-C, Tables III and IV] The new-class test sets are generated by the same procedural pipeline as the training sets: Section III-B describes the manually authored InfraredEffect shader nodes, Section III-C describes the Cycles rendering and Alpha Over compositing, and Section IV-A specifies Nconfig=2 for HIT-UAV and Nconfig=1 for MONET, with each test image containing exactly one synthetic drone (1,161 instances per 1,158 images) or one synthetic deer (7,224 instances per 7,224 images). Because the same 3D models, shaders, renderer, compositing routine, and annotation-projection code are used for both training and test images, the reported mAP@0.5 values of 0.995 and 0.930 measure intra-pipeline consistency rather than detection of real drones or deer in thermal imagery. The detector can exploit pipeline-specific cues such as perfect object silhouettes, consistent thermal gradients, and compositing edges. No real thermal image of a drone or deer is used in training or evaluation, so the abstract's claim that these results 'validate the successful expansion into new applications' is not supported. Section V discloses the need for manually modeled thermal shading but does not identify this evaluation gap.
  2. [Section IV-D, Table V] The comparison between the DUT-Anti-UAV visible-light detector and the HIT-UAV-drone thermal detector is confounded by multiple factors beyond sensor modality: the training images differ in camera perspective (ground-level versus aerial), scene content, object scale, and annotation protocol. Evaluating a visible-light model on thermal test images is expected to produce low mAP because of the modality shift alone, so this experiment cannot support the paper's claim that 'thermal detectors outperform their visible-light-trained counterparts' or the broader conclusion about the domain gap. An apples-to-apples comparison would require training on visible and thermal data captured under matched conditions, ideally with the same camera positions, scenes, and object instances.
minor comments (5)
  1. [Title and Section I] The text contains the typo 'UA V' in the title and in the first paragraph; it should read 'UAV'.
  2. [Section III-B] The parameter values for the thermal shader nodes (emissivity, Color Ramp thresholds, Fresnel settings) are not reported; providing these values or a node graph would improve reproducibility and would make the sensitivity of the results to these choices assessable.
  3. [Section IV-C and Section I] The paper lists oriented bounding boxes and segmentation masks as contributions, but no experiment evaluates the quality of these annotations; even a small manually labeled subset with IoU for OBB and masks would be helpful.
  4. [Tables III and IV] Because every test image contains exactly one new-class instance, the new-class precision/recall and mAP values may be optimistic; future evaluations should include test images without the new class and images with multiple instances to better reflect real deployment conditions.
  5. [Section IV-D, Table VII] The 'Random [0,90]' and 'Metadata aligned' mAP values are very close (0.981 vs. 0.995); the paper should report whether this difference is stable across multiple training runs or is within run-to-run variance.

Circularity Check

1 steps flagged · score 6.0 of 10

New-class mAP values are measured on test images generated by the same procedural pipeline as the training images, so the headline numbers measure intra-pipeline consistency rather than transfer to real thermal data.

  1. self definitional [Section IV-C (Tables III and IV) with Section IV-B dataset construction]
    "For the HIT-UAV-drone dataset, we incorporate the 'drone' class, which is crucial for urban mobility and safety. Similar to the MONET-deer dataset, the new class is synthetically generated and integrated with existing thermal data. To further enhance the dataset, we double its size by generating two virtual scenes for each background. ... In the table Table III we presents results of training on the HIT-UAV-drone dataset and evaluating on the test split of the dataset."

    For both new classes, the training and test splits are produced by the same procedural pipeline: the same 3D models, the same manually authored InfraredEffect shader nodes, the same Cycles renderer, the same Alpha Over compositing routine, and the same annotation-projection code. The HIT-UAV-drone test set is literally derived from the same backgrounds and same two-virtual-scenes-per-background generation process described in Section IV-A. Thus the reported drone mAP@0.5 = 0.995 and deer mAP@0.5 = 0.930 are consistency scores between two samples drawn from the same generative distribution, not measures of out-of-distribution transfer to real thermal imagery of drones or deer.

full rationale

The paper makes a genuine methodological contribution: a procedural Blender pipeline for compositing synthetic objects onto real thermal backgrounds, with automatic AABB/OBB/mask annotations. The existing-class results (Tables III, IV, VI) and the aerial-perspective ablation (Table VII) provide partially independent evidence, since the original classes' test instances come from the real HIT-UAV and MONET datasets. However, the headline new-class performance is circular in the following specific sense: the new-class test images are generated by the same pipeline as the new-class training images, sharing the same 3D mesh, material shader, renderer, and overlay code. No real thermal image of a drone or deer is used in training or evaluation for those classes. Consequently, the mAP values of 0.995 and 0.930 do not demonstrate that the synthetic appearances are realistic enough to transfer to real LWIR data; they demonstrate only that a detector can recognize the pipeline's own output. The limitations section acknowledges that manual thermal shading is required for each object but does not disclose this evaluation gap. This is a partial circularity, concentrated in the central new-class claim, so a score of 6 is appropriate.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No new physical or mathematical entities are introduced. The central claims rest on modeling assumptions about thermal realism and viewpoint transfer, plus manual shader parameters; because evaluation shares the generation pipeline, these assumptions are not independently verified.

free parameters (1)
  • Thermal material shader parameters (emissivity, Color Ramp thresholds, Fresnel settings) for drone and deer models = not reported
    Section III-B states the InfraredEffect node is manually modeled for each object; the realism of the synthetic thermal appearance depends on hand-set values that are not specified.
assumptions (4)
  • domain assumption Blender Cycles with custom emission and Fresnel nodes realistically approximates real thermal appearance for synthetic objects
    Section III-B describes thermal shader modeling but provides no calibration, radiometric data, or comparison with real thermal measurements.
  • domain assumption Metadata camera angles in HIT-UAV and MONET correctly describe capture geometry so that virtual viewpoint alignment matches backgrounds
    Section IV-A relies on metadata; the authors exclude subsets with missing metadata and note negative gimbal pitch values, so this assumption is only partially satisfied.
  • ad hoc to paper Objects rendered in Blender can be overlaid on real thermal backgrounds without creating a harmful domain gap for detector training
    Section III-C uses Alpha Over compositing; the paper's validity depends on this overlay realism, but no real-world test is provided.
  • ad hoc to paper Test images produced by the same synthetic pipeline are representative of real thermal aerial images of the new classes
    Section IV-C evaluates on synthetic test splits; this assumption is untested and is the key gap between the reported mAP and the claimed applications.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unlocking Thermal Aerial Imaging: Synthetic Enhancement of UAV Datasets." pith.science (2026). https://pith.science/paper/JGYC5EOM

@misc{pith2026250706797,
  author       = {Pith},
  title        = {Pith review of: Unlocking Thermal Aerial Imaging: Synthetic Enhancement of UAV Datasets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JGYC5EOM}},
  note         = {Machine review of arXiv:2507.06797}
}
read the original abstract

Thermal imaging from unmanned aerial vehicles (UAVs) holds significant potential for applications in search and rescue, wildlife monitoring, and emergency response, especially under low-light or obscured conditions. However, the scarcity of large-scale, diverse thermal aerial datasets limits the advancement of deep learning models in this domain, primarily due to the high cost and logistical challenges of collecting thermal data. In this work, we introduce a novel procedural pipeline for generating synthetic thermal images from an aerial perspective. Our method integrates arbitrary object classes into existing thermal backgrounds by providing control over the position, scale, and orientation of the new objects, while aligning them with the viewpoints of the background. We enhance existing thermal datasets by introducing new object categories, specifically adding a drone class in urban environments to the HIT-UAV dataset and an animal category to the MONET dataset. In evaluating these datasets for object detection task, we showcase strong performance across both new and existing classes, validating the successful expansion into new applications. Through comparative analysis, we show that thermal detectors outperform their visible-light-trained counterparts and highlight the importance of replicating aerial viewing angles. Project page: https://github.com/larics/thermal_aerial_synthetic.

Figures

Figures reproduced from arXiv: 2507.06797 by the authors.

Figure 1
Figure 1. Procedural pipeline for generating synthetic thermal images. Starting from the original HIT-UAV dataset, the thermal background and the corresponding annotations are used as the base. A new virtual scene is procedurally generated, based on configurable parameters, while aligning the camera view with the background’s perspective. After rendering new objects, they are overlaid onto the thermal background. Annotations … view at source ↗
Figure 2
Figure 2. 3D thermal models used in the procedural pipeline for synthetic [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Annotation types in the HIT-UAV-drone dataset. We merged [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Example data from the synthetically enhanced datasets, showcasing varied viewpoints recreated in a virtual environment using metadata alignment. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages

  1. [1]

    A review on deep learning in uav remote sensing,

    L. P. Osco, J. Marcato Junior, A. P. Marques Ramos, L. A. de Cas- tro Jorge, S. N. Fatholahi, J. de Andrade Silva, E. T. Matsubara, H. Pis- tori, W. N. Gonc ¸alves, and J. Li, “A review on deep learning in uav remote sensing,” International Journal of Applied Earth Observation and Geoinformation, vol. 102, p. 102456, Oct. 2021

    work page 2021

  2. [2]

    Application of raycast method for person geolocalization and distance determination using uav images in real-world land search and rescue scenarios,

    G. Paulin, S. Sambolek, and M. Ivasic-Kos, “Application of raycast method for person geolocalization and distance determination using uav images in real-world land search and rescue scenarios,” Expert Systems with Applications , vol. 237, p. 121495, Mar. 2024

    work page 2024

  3. [3]

    Tvub: Thermal vision-based uav and blockchain-aided poaching prevention system,

    S. Anbalagan, W. Alhakami, A. Manoharan, S. G. S, A. Nehra, H. Alhakami, and G. Raja, “Tvub: Thermal vision-based uav and blockchain-aided poaching prevention system,” in GLOBECOM 2023 - 2023 IEEE Global Communications Conference , IEEE, Dec. 2023

    work page 2023

  4. [4]

    Nonsalient object detection algorithm based on infrared image,

    Z. Ye and Z. Cai, “Nonsalient object detection algorithm based on infrared image,” IEEE Geoscience and Remote Sensing Letters , vol. 20, pp. 1–5, 2023

    work page 2023

  5. [5]

    Unmanned aerial sensor placement for cluttered environments,

    A. Farinha, R. Zufferey, P. Zheng, S. F. Armanini, and M. Kovac, “Unmanned aerial sensor placement for cluttered environments,” IEEE Robotics and Automation Letters , vol. 5, p. 6623–6630, Oct. 2020

    work page 2020

  6. [6]

    Drone-assisted collection of environmental dna from tree branches for biodiversity monitoring,

    E. Aucone, S. Kirchgeorg, A. Valentini, L. Pellissier, K. Deiner, and S. Mintchev, “Drone-assisted collection of environmental dna from tree branches for biodiversity monitoring,” Science Robotics , vol. 8, Jan. 2023

    work page 2023

  7. [7]

    People detection and tracking from aerial thermal views,

    J. Portmann, S. Lynen, M. Chli, and R. Siegwart, “People detection and tracking from aerial thermal views,” in 2014 IEEE International Conference on Robotics and Automation (ICRA) , pp. 1794–1800, 2014

    work page 2014

  8. [8]

    Birdsai: A dataset for detection and tracking in aerial thermal infrared videos,

    E. Bondi, R. Jain, P. Aggrawal, S. Anand, R. Hannaford, A. Kapoor, J. Piavis, S. Shah, L. Joppa, B. Dilkina, and M. Tambe, “Birdsai: A dataset for detection and tracking in aerial thermal infrared videos,” in 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 1736–1745, 2020

    work page 2020

Show all 18 references
  1. [9]

    Wisard: A labeled visual and thermal image dataset for wilderness search and rescue,

    D. Broyles, C. R. Hayner, and K. Leung, “Wisard: A labeled visual and thermal image dataset for wilderness search and rescue,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE, Oct. 2022

    2022

  2. [10]

    Hit-uav: A high-altitude infrared thermal dataset for unmanned aerial vehicle- based object detection,

    J. Suo, T. Wang, X. Zhang, H. Chen, W. Zhou, and W. Shi, “Hit-uav: A high-altitude infrared thermal dataset for unmanned aerial vehicle- based object detection,” Scientific Data, vol. 10, Apr. 2023

    2023

  3. [11]

    The monet dataset: Multimodal drone thermal dataset recorded in rural scenarios,

    L. Riz, A. Caraffa, M. Bortolon, M. L. Mekhalfi, D. Boscani, A. Moura, J. Antunes, A. Dias, H. Silva, A. Leonidou, C. Constan- tinides, C. Keleshis, D. Abate, and F. Poiesi, “The monet dataset: Multimodal drone thermal dataset recorded in rural scenarios,” inIEEE Proc. of the ...

    2023

  4. [12]

    Wit-uas: A wildland-fire infrared thermal dataset to detect crew assets from aerial views,

    A. Jong, M. Yu, D. Dhrafani, S. Kailas, B. Moon, K. Sycara, and S. Scherer, “Wit-uas: A wildland-fire infrared thermal dataset to detect crew assets from aerial views,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pp. 11464– 11471, 2023

    2023

  5. [13]

    Sim2air - synthetic aerial dataset for uav monitoring,

    A. Barisic, F. Petric, and S. Bogdan, “Sim2air - synthetic aerial dataset for uav monitoring,” IEEE Robotics and Automation Letters , vol. 7, p. 3757–3764, Apr. 2022

    2022

  6. [14]

    Blenderproc2: A procedural pipeline for photorealistic rendering,

    M. Denninger, D. Winkelbauer, M. Sundermeyer, W. Boerdijk, M. Knauer, K. H. Strobl, M. Humt, and R. Triebel, “Blenderproc2: A procedural pipeline for photorealistic rendering,” Journal of Open Source Software, vol. 8, no. 82, p. 4901, 2023

    2023

  7. [15]

    Airsim-w: A simula- tion environment for wildlife conservation with uavs,

    E. Bondi, D. Dey, A. Kapoor, J. Piavis, S. Shah, F. Fang, B. Dilkina, R. Hannaford, A. Iyer, L. Joppa, and M. Tambe, “Airsim-w: A simula- tion environment for wildlife conservation with uavs,” in Proceedings of the 1st ACM SIGCAS Conference on Computing and Sustainable Societi...

    2018

  8. [16]

    Thermalsynth: A novel approach for generating synthetic thermal human scenarios,

    N. Madan, M. S. N. Siemon, M. K. Gjerde, B. S. Petersson, A. Gro- tuzas, M. A. Esbensen, I. A. Nikolov, M. P. Philipsen, K. Nasrollahi, and T. B. Moeslund, “Thermalsynth: A novel approach for generating synthetic thermal human scenarios,” in 2023 IEEE/CVF Winter Con- ference o...

    2023

  9. [17]

    Bio- inspired visual relative localization for large swarms of uavs,

    M. K ˇr´ıˇzek, M. Vrba, A. B. Kula ˇs, S. Bogdan, and M. Saska, “Bio- inspired visual relative localization for large swarms of uavs,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) , vol. 16, p. 11825–11831, IEEE, May 2024

    2024

  10. [18]

    Blender - a 3d modelling and rendering package,

    Blender Online Community, “Blender - a 3d modelling and rendering package,” 2024. Blender Foundation, Stichting Blender Foundation, Amsterdam

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.