Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Model-Agnostic Open-Set Air-to-Air Visual Object Detection for Reliable UAV Perception

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read By fusing softmax confidence with embedding-space Gaussian mixture entropy, this paper proposes a post-hoc open-set detection layer that lets real-time aerial detectors reject unseen objects without retraining, reporting that it raises OOD

desk verdict The paper builds a plausible UAV open-set detector from known components, but the main AUROC claim rests on an undefined score for Joint Thresholding and on selective reporting. read the letter →

arxiv 2509.09297 v1 pith:MHJXRWHF submitted 2025-09-11 cs.CV cs.LGcs.RO

Spyridon Loukovitis , Anastasios Arsenos , Vasileios Karampinis , Athanasios Voulodimos This is my paper · ORCID
classification cs.CVcs.LGcs.RO
keywords open-setobjectdetectionUAVperceptionout-of-distributionGaussianmixturemodelspectralnormalizationair-to-airAOT-Cbenchmarkreal-time
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 tries to show that an open-set detection layer can be added to any embedding-based object detector, making UAV perception able to flag unknown aerial objects instead of silently misclassifying them. The layer estimates semantic uncertainty by fitting per-class Gaussian mixtures in embedding space, combines that with softmax confidence in a joint threshold, and regularizes the backbone with spectral normalization. The core claim is that this fusion is more robust to real-world flight conditions than any single confidence score, including plain softmax or GMM density. On the AOT-C benchmark and real flight data, the method reports AUROC improvements from 0.800 to 0.883/0.887 compared to standard YOLOv5, with no significant loss in detection mAP and negligible impact on inference speed. If this holds, UAVs could reject novel threats and corrupted inputs without architectural changes or retraining.

What carries the argument

The carrying mechanism is Joint Thresholding: each detection is kept only if its softmax confidence exceeds a threshold AND its GMM posterior entropy falls below a threshold. The GMMs are fit post-training on training embeddings matched to ground truth; spectral normalization on the backbone enforces a bi-Lipschitz constraint for well-behaved embeddings, and temperature scaling learned on validation recalibrates both logits and GMM log-likelihoods. The key unit is the per-detection embedding, which lets the same machinery work with any detector that outputs fixed-dimensional embeddings.

What would settle it

Re-run the Joint Thresholding pipeline with thresholds fixed using only the AOT-C training/validation splits, then evaluate on an independently collected aerial flight dataset that was never used to select the configuration; if the AUROC_bd gap over YOLOv5 (reported 0.883/0.887 vs 0.800) shrinks to near zero or reverses, the central claim would be falsified. A simpler check: repeat the real-flight evaluation with the softmax-entropy variant excluded from the best-configuration selection and verify the gap persists.

Watch

Extended reading notes

Core claim

The central discovery is that softmax-derived confidence and embedding-space GMM entropy carry complementary signals about out-of-distribution inputs, and fusing them via a simple threshold—Joint Thresholding—separates known aircraft from unknown drones better than any single score. The authors show that on real flight data, after training on AOT-C, Joint Thresholding reaches AUROC_bd of 0.883 (RT-DETR) and 0.887 (with spectral normalization), versus 0.800 for YOLOv5, while maintaining mAP around 40. The method also treats background detections as OOD and improves rejection there. This is presented as a post-hoc calibration layer that requires no retraining or detector modification, only acc

Load-bearing premise

The load-bearing premise is that the best-performing configuration found on the AOT validation OOD set (including the joint-threshold parameters) generalizes to new flight data without having been tuned on that data.

Editorial extensions

If this is right

  • Any embedding-based detector (not just RT-DETR) can be retrofitted with open-set rejection by fitting GMMs on training embeddings and applying joint thresholding; no architecture change or retraining is required.
  • Treating background detections as OOD rather than ignoring them improves AUROC and better reflects real UAV false-positive pressure, so background rejection becomes a trainable part of the open-set pipeline.
  • Spectral normalization and temperature scaling are effective post-hoc calibrations: SN consistently raises AUROC across scoring methods, while temperature scaling is needed especially for GMM-based scores.
  • Real-time operation is preserved; the method runs at about 24 FPS on an A10G, effectively identical to baseline RT-DETR.
  • Softmax entropy alone, despite strong synthetic results, degrades on real flight data, so multi-signal fusion is necessary for deployment.

Reading between the lines

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

  • The paper leaves implicit that its fixed-threshold fusion is a special case of score ensembling; a learned fusor (e.g., decision tree or shallow MLP over the same score vector) might close the remaining gap and would be directly testable with the released code.
  • A natural extension the authors do not test is whether the same GMM-entropy layer transfers to other small-object aerial benchmarks (e.g., Det-Fly) or to different backbone architectures; if it does, the 'model-agnostic' claim would be stronger than the single RT-DETR demonstration.
  • Because thresholds are tuned on a validation OOD split containing drones, operational deployment would need a principled way to set thresholds without such a split; an unsupervised entropy calibration or per-mission threshold adaptation might work, but the paper does not address this.
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

4 major / 5 minor

Summary. The paper proposes a model-agnostic open-set detection framework for air-to-air UAV perception. It augments an embedding-based detector (RT-DETR-R50) with per-detection GMM density estimation, spectral normalization, temperature scaling, and a Joint Thresholding rule that combines softmax confidence and GMM entropy to reject OOD detections. Experiments on AOT-C and real flight data report AUROC gains over YOLOv5 and GMM-Det while preserving mAP and real-time speed. The central claims are that joint thresholding improves open-set rejection under corruption and domain shift while introducing negligible runtime overhead.

Significance. The paper addresses a safety-relevant problem in aerial robotics and gives a sensible combination of existing post-hoc uncertainty techniques (DDU, GMM-Det, temperature scaling) for small moving aerial targets. Contributions include a broad ablation over softmax/entropy/GMM scoring functions, investigation of spectral normalization, and an evaluation on real flight data rather than synthetic corruptions only. The authors also disclose their configuration-selection protocol, which is helpful. If the quantitative claims held, the method would be practically useful. However, as presented, the central quantitative claim is not verifiable because the primary metric (AUROC) is undefined for the proposed hard-decision rule, and the experimental reporting does not permit a fair comparison. The work is promising but requires substantial methodological clarification and additional experiments.

major comments (4)
  1. [Section III-D, Algorithm 1; Tables II, IV; Fig. 4] Joint Thresholding is defined as a hard decision rule: a detection is ID only if s_softmax >= tau_soft AND H_gmm <= tau_gmm, otherwise OOD. This yields a binary label, not a continuous ranking score. AUROC and ROC require a score over all thresholds, but the paper never defines such a score for this conjunction rule. The reported AUROC values (0.982 in Table II, 0.883-0.887 in Table IV) and Figure 4 are therefore not reproducible. Define a single continuous score (e.g., a monotone function of s_softmax and -H_gmm) that reduces to the stated rule, or replace AUROC with threshold-dependent metrics at specified operating points.
  2. [Section IV-C] The paper states: 'Each method is evaluated across multiple configurations ... and only the best-performing setup is shown. Results without softmax pruning are omitted.' This is post-hoc selection over a large configuration space (embedding layer, GMM components, temperatures, pruning threshold, spectral normalization). Without reporting all configurations or a pre-specified validation selection protocol, Tables II and IV do not support a fair comparison. No error bars or multiple seeds are reported, so the 0.982 vs. 0.929 gap may be within noise. Report full distributions, the split used for selection, and variability across seeds.
  3. [Section IV-C/D, Tables III-IV] The closed-set mAP claim is confounded: Table III reports mAP only for the best-AUROC configuration per method, with no comparison to the unfiltered detector or to non-best configurations. In Table IV, configurations are selected on AOT-C validation and the real-flight images come from the authors' prior benchmark [5], which weakens the transfer claim. Report mAP for all filtering modes and clarify that no test information was used in configuration selection; also consider an external test split.
  4. [Table IV] The headline gain versus YOLOv5 is not an apples-to-apples comparison of the proposed open-set method: YOLOv5 is evaluated with its standard inference, without the paper's embedding-GMM or pruning pipeline. Instead of comparing RT-DETR+Joint Thresholding to standard YOLOv5, the reader needs YOLOv5 baseline vs. YOLOv5 with the same open-set pipeline to support the model-agnostic claim. Apply the framework to YOLOv5 or another detector and report the relative gain with that detector's own baseline.
minor comments (5)
  1. [Section III-C] The pruning threshold S_max < 0.2 is introduced as a fixed choice, but no sensitivity analysis is given. A brief study over this threshold would clarify whether results depend strongly on it.
  2. [Table II] The header 'TPR@OSR 5%/10%/20%' uses 'OSR' without definition. Presumably it denotes a fixed open-set rate or false-positive rate; state this explicitly.
  3. [Algorithm 1] The notation in lines 9-11 is confusing: 'x_c' is used for both an image and a set of samples. Use clearer notation such as X_c for the set of matched embeddings.
  4. [Figure 3 caption] The caption says 'same image' but shows left/right panels from different detectors; clarify whether the panels show different views or crops of the same scene.
  5. [General] Typos and formatting issues: 'UA V' contains a spurious space; Table I headers use 'APclean' and 'APcor' without separation; reference [28] is an informal GitHub-style citation.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the reported AUROC is an empirical measurement, not a quantity forced by the paper's definitions; the main caveats (undefined ranking score for Joint Thresholding, best-config selection, self-created benchmark) are correctness/independence concerns rather than circularity.

full rationale

The claimed derivation chain is not circular in the sense defined by the review criteria. The GMM density model is fitted on in-distribution training embeddings only; the paper explicitly states 'No OOD data are used at this stage' (Section III-B2), so the OOD separation is not inserted by construction. Spectral normalization and temperature scaling are imported from external work ([3], [4]) and applied post hoc; the paper does not derive its headline result from these citations. Joint Thresholding is an empirical conjunction of a softmax score and a GMM-entropy threshold, and its reported AUROC is measured on data, not obtained by a definitional identity. The most serious concerns are: (i) Algorithm 1 returns a binary ID/OOD decision, yet Tables II and IV report AUROC/ROC curves, so a continuous ranking score for Joint Thresholding is never defined; (ii) Section IV-C states 'Each method is evaluated across multiple configurations ... and only the best-performing setup is shown', which implies configuration selection on the same OOD set whose AUROC is then reported — a statistical overfitting risk, but not a derivation equivalent to the inputs; (iii) the AOT-C and real-flight test data come from the authors' prior work [5], which is a dataset-independence caveat, not a circular step. These issues bear on reproducibility and validity more than on circularity. Accordingly, the circularity score is low.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities or conserved quantities. The central claim rests on the empirical regularity of the embedding space, on the transfer from the AOT-C corruption benchmark to real flight, and on the quality of the real-flight labels. The free parameters are thresholds, temperatures, GMM component counts, and the embedding layer, all selected empirically during evaluation. These choices are not reported exhaustively, which weakens the independence of the headline AUROC numbers.

free parameters (8)
  • tau_softmax
    Threshold for softmax score in Joint Thresholding (Algorithm 1); selected on validation set; numerical value not reported.
  • tau_gmm
    Threshold for GMM entropy in Joint Thresholding (Algorithm 1); selected on validation set; numerical value not reported.
  • T_model
    Temperature for softmax logits, learned on validation via negative-log-likelihood minimization (Section III-C2).
  • T_gmm
    Temperature for GMM log-likelihoods, learned on validation via NLL minimization (Section III-C2).
  • GMM components K per class = 2, 3, 4
    Number of Gaussians per class in the Multi-GMM variant, chosen empirically (Section III-B2).
  • embedding layer choice
    Which detector layer provides the embeddings; the best layer is chosen per method during the ablation (Table II caption: 'only the best-performing setup is shown').
  • softmax pruning threshold = 0.2
    Fixed threshold below which detections are discarded (Section III-C1); value is stated but not tuned.
  • spectral normalization on/off = bool
    Binary choice of whether to apply spectral normalization to the backbone; evaluated as a variant rather than a fitted continuous parameter.
assumptions (5)
  • domain assumption The detector's embedding space is regular enough for per-class Gaussian models to separate in-distribution from out-of-distribution objects.
    The entire GMM-based OOD scoring in Section III-B depends on this; DDU-style bi-Lipschitz regularization is invoked as a helper.
  • domain assumption The AOT-C corruption model is a representative proxy for real-world flight conditions.
    The method is trained on AOT-C corruptions and then evaluated on real flight data (Section IV-D); transfer from synthetic corruptions to reality is assumed.
  • domain assumption The real-flight test set has reliable ID/OOD labels and a known object distribution.
    The paper does not describe the real-flight dataset composition (object types, ID/OOD mix, background fraction). AUROC_bd treats background as OOD, which is only meaningful if background detections are identifiable as errors.
  • domain assumption Class-conditional GMMs fitted on training embeddings generalize to unseen OOD samples without using OOD data for fitting.
    Standard OOD detection assumption; stated in Section III-B: 'No OOD data are used at this stage.'
  • domain assumption Temperature scaling learned on the validation set transfers to the test distribution.
    Post-hoc calibration (Section III-C2) is assumed to improve separation on held-out data; if the validation distribution differs from the OOD distribution, the benefit may vanish.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Model-Agnostic Open-Set Air-to-Air Visual Object Detection for Reliable UAV Perception." pith.science (2026). https://pith.science/paper/MHJXRWHF

@misc{pith2026250909297,
  author       = {Pith},
  title        = {Pith review of: Model-Agnostic Open-Set Air-to-Air Visual Object Detection for Reliable UAV Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MHJXRWHF}},
  note         = {Machine review of arXiv:2509.09297}
}
read the original abstract

Open-set detection is crucial for robust UAV autonomy in air-to-air object detection under real-world conditions. Traditional closed-set detectors degrade significantly under domain shifts and flight data corruption, posing risks to safety-critical applications. We propose a novel, model-agnostic open-set detection framework designed specifically for embedding-based detectors. The method explicitly handles unknown object rejection while maintaining robustness against corrupted flight data. It estimates semantic uncertainty via entropy modeling in the embedding space and incorporates spectral normalization and temperature scaling to enhance open-set discrimination. We validate our approach on the challenging AOT aerial benchmark and through extensive real-world flight tests. Comprehensive ablation studies demonstrate consistent improvements over baseline methods, achieving up to a 10\% relative AUROC gain compared to standard YOLO-based detectors. Additionally, we show that background rejection further strengthens robustness without compromising detection accuracy, making our solution particularly well-suited for reliable UAV perception in dynamic air-to-air environments.

Figures

Figures reproduced from arXiv: 2509.09297 by the authors.

Figure 1
Figure 1. Overview of the object detection and uncertainty estimation pipeline. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Distribution of softmax scores for in-distribution (blue) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Side-by-side comparison for the same image: the left half of every panel shows RT-DETR (SN), the right half shows YOLO. Top row contains in-distribution (ID) objects, while the bottom row contains out-of-distribution (OOD/ID) objects. A blue box indicates the detector classified the object as ID; a red box indicates the detector judged it OOD. RT-DETR correctly classifies the planes (ID) and the drones (OOD) in all … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison of ROC curves for different methods in open-set real flight data. (a) Results ignoring background detections. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 3 linked inside Pith

  1. [5]

    Common corruptions for evaluating and enhancing robustness in air-to-air visual object detection,

    A. Arsenos, V . Karampinis, E. Petrongonas, C. Skliros, D. Kollias, S. Kollias, and A. V oulodimos, “Common corruptions for evaluating and enhancing robustness in air-to-air visual object detection,”IEEE Robotics and Automation Letters, vol. 9, no. 7, pp. 6688–6695, 2024

    2024

  2. [1]

    Class semantics modulation for open-set instance segmentation,

    Y . Yang, Z. Zhou, J. Wu, Y . Wang, and R. Xiong, “Class semantics modulation for open-set instance segmentation,”IEEE Robotics and Automation Letters, vol. 9, no. 3, pp. 2240–2247, 2024

    2024

  3. [2]

    Open-set object detection using classification-free object proposal and instance-level contrastive learning,

    Z. Zhou, Y . Yang, Y . Wang, and R. Xiong, “Open-set object detection using classification-free object proposal and instance-level contrastive learning,”IEEE Robotics and Automation Letters, vol. 8, no. 3, pp. 1691–1698, 2023

    2023

  4. [3]

    Deep deterministic uncertainty: A new simple baseline,

    J. Mukhoti, A. Kirsch, J. Van Amersfoort, P. H. Torr, and Y . Gal, “Deep deterministic uncertainty: A new simple baseline,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 24 384–24 394

    2023

  5. [4]

    Uncertainty for identifying open-set errors in visual object detection,

    D. Miller, N. S ¨underhauf, M. Milford, and F. Dayoub, “Uncertainty for identifying open-set errors in visual object detection,”IEEE Robotics and Automation Letters, vol. 7, no. 1, pp. 215–222, 2021

    2021

  6. [6]

    Air-to-air visual detection of micro-uavs: An experimental evaluation of deep learning,

    Y . Zheng, Z. Chen, D. Lv, Z. Li, Z. Lan, and S. Zhao, “Air-to-air visual detection of micro-uavs: An experimental evaluation of deep learning,” IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 1020–1027, 2021

    2021

  7. [7]

    Nefeli: A deep-learning detection and tracking pipeline for enhancing autonomy in advanced air mobility,

    A. Arsenos, E. Petrongonas, O. Filippopoulos, C. Skliros, D. Kollias, and S. Kollias, “Nefeli: A deep-learning detection and tracking pipeline for enhancing autonomy in advanced air mobility,”Aerospace Science and Technology, vol. 155, p. 109613, 2024

    2024

  8. [8]

    Airtrack: Onboard deep learning framework for long-range aircraft detection and tracking,

    S. Ghosh, J. Patrikar, B. Moon, M. M. Hamidi, and S. Scherer, “Airtrack: Onboard deep learning framework for long-range aircraft detection and tracking,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 1277–1283

    2023

Show all 34 references
  1. [9]

    Visual-based obstacle detection and tracking, and conflict detection for small uas sense and avoid,

    R. Opromolla and G. Fasano, “Visual-based obstacle detection and tracking, and conflict detection for small uas sense and avoid,”Aerospace Science and Technology, vol. 119, p. 107167, 2021. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1270963821006775

    2021

  2. [10]

    Out-of-distribution identification: Let detector tell which i am not sure,

    R. Li, C. Zhang, H. Zhou, C. Shi, and Y . Luo, “Out-of-distribution identification: Let detector tell which i am not sure,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 638–654

    2022

  3. [11]

    Hyperdimensional feature fusion for out-of-distribution detection,

    S. Wilson, T. Fischer, N. S ¨underhauf, and F. Dayoub, “Hyperdimensional feature fusion for out-of-distribution detection,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023, pp. 2644–2654

    2023

  4. [12]

    Open-set object detection: towards unified problem formulation and benchmarking,

    H. Ammar, N. Kiselov, G. Lapouge, and R. Audigier, “Open-set object detection: towards unified problem formulation and benchmarking,” in European Conference on Computer Vision. Springer, 2024, pp. 46–61

    2024

  5. [13]

    Grounding dino 1.5: Advance the

    T. Ren, Q. Jiang, S. Liu, Z. Zeng, W. Liu, H. Gao, H. Huang, Z. Ma, X. Jiang, Y . Chenet al., “Grounding dino 1.5: Advance the” edge” of open-set object detection,”arXiv preprint arXiv:2405.10300, 2024

    2024 arXiv

  6. [14]

    Uncertainty quantification for safe and reliable autonomous vehicles: A review of methods and applications,

    K. Wang, C. Shen, X. Li, and J. Lu, “Uncertainty quantification for safe and reliable autonomous vehicles: A review of methods and applications,”IEEE Transactions on Intelligent Transportation Systems, 2025

    2025

  7. [15]

    Gaussian mixture models,

    D. Reynolds, “Gaussian mixture models,” inEncyclopedia of biometrics. Springer, 2015, pp. 827–832

    2015

  8. [16]

    Certainnet: Sampling-free uncertainty estimation for object detection,

    S. Gasperini, J. Haug, M.-A. N. Mahani, A. Marcos-Ramiro, N. Navab, B. Busam, and F. Tombari, “Certainnet: Sampling-free uncertainty estimation for object detection,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 698–705, 2021

    2021

  9. [17]

    Spectral normal- ization for generative adversarial networks,

    T. Miyato, T. Kataoka, M. Koyama, and Y . Yoshida, “Spectral normal- ization for generative adversarial networks,” inInternational Conference on Learning Representations (ICLR), 2018

    2018

  10. [18]

    On calibration of modern neural networks,

    C. Guo, G. Pleiss, Y . Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” inProceedings of the 34th International Conference on Machine Learning (ICML), vol. 70, 2017, pp. 1321–1330

    2017

  11. [19]

    Benchmarking neural network ro- bustness to common corruptions and perturbations,

    D. Hendrycks and T. Dietterich, “Benchmarking neural network ro- bustness to common corruptions and perturbations,”Proceedings of the International Conference on Learning Representations, 2019

    2019

  12. [20]

    The cityscapes dataset for semantic urban scene understanding,

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benen- son, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” inProc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  13. [21]

    nuscenes: A multimodal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for autonomous driving,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11 621–11 631

    2020

  14. [22]

    ultralytics/yolov5: v3.1 - bug fixes and performance improvements,

    G. Jocher, A. Stoken, J. Borovec, NanoCode012, ChristopherSTAN, L. Changyu, Laughing, tkianai, A. Hogan, lorenzomammana, yxNONG, AlexWang1900, L. Diaconu, Marc, wanghaoyang0106, ml5ah, Doug, F. Ingham, Frederik, Guilhen, Hatovix, J. Poznanski, J. Fang, L. Yu, changyu98, M. Wan...

    2020 doi

  15. [23]

    Ultralytics yolo,

    G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics yolo,” https://github. com/ultralytics/ultralytics, Jan 2023, [Online; accessed August 2, 2025]

    2023

  16. [24]

    Yolox: Exceeding yolo series in 2021,

    Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, “Yolox: Exceeding yolo series in 2021,”arXiv preprint arXiv:2107.08430, 2021

    2021 arXiv

  17. [25]

    Focal loss for dense object detection,

    T. Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” inProceedings of the IEEE International Conference on Computer Vision (ICCV), 2017

    2017

  18. [26]

    Pytorch-retinanet: Pytorch implementation of retinanet,

    Y . Henon, “Pytorch-retinanet: Pytorch implementation of retinanet,”

  19. [27]

    Faster r-cnn: Towards real-time object detection with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 28, 2015

    2015

  20. [28]

    Faster r-cnn pytorch training pipeline,

    sovit-123, “Faster r-cnn pytorch training pipeline,” 2025. [Online]. Avail- able: https://github.com/sovit-123/fasterrcnn-pytorch-training-pipeline

    2025

  21. [29]

    Diffusiondet: Diffusion model for object detection,

    S. Chen, P. Sun, Y . Song, and P. Luo, “Diffusiondet: Diffusion model for object detection,” in2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 19 773–19 786

    2023

  22. [30]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in European conference on computer vision. Springer, 2020, pp. 213– 229

    2020

  23. [31]

    Probabilistic two-stage detec- tion,

    X. Zhou, V . Koltun, and P. Kr ¨ahenb¨uhl, “Probabilistic two-stage detec- tion,” inarXiv preprint arXiv:2103.07461, 2021

    2021 arXiv

  24. [32]

    DETRs Beat YOLOs on Real-time Object Detection,

    Y . Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y . Liu, and J. Chen, “DETRs Beat YOLOs on Real-time Object Detection,” Seattle, Washington, USA, pp. 16 965–16 974, June 2024

    2024

  25. [33]

    Airborne object tracking dataset,

    “Airborne object tracking dataset,” https://registry.opendata.aws/ airborne-object-tracking, accessed: 2023-07-23

    2023

  26. [2020]

    Available: https://github.com/yhenon/pytorchretinanet

    [Online]. Available: https://github.com/yhenon/pytorchretinanet

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.