Pith. sign in

REVIEW 3 major objections 5 minor 55 references

Enhancing people localisation in drone imagery for better crowd management by utilising every pixel in high-resolution images

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Drone crowd localisation reaches a new state of the art by processing every pixel of high-resolution images rather than downscaling or sliding a window.

desk verdict A solid engineering paper with a genuinely useful dataset; the SOTA claim is undercut by a quoted baseline and an internal inconsistency, but both are fixable. read the letter →

arxiv 2502.04014 v1 pith:BPYN5TX5 submitted 2025-02-06 cs.CV cs.RO

Bartosz Ptak , Marek Kraft This is my paper · ORCID
classification cs.CVcs.RO
keywords peoplelocalisationdroneimagerycrowdmanagementhigh-resolutionimageprocessingPixelDistillpoint-orientedobjectaerialremotesensingUP-COUNTdataset
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 claims that people in high-resolution drone images can be localised as points more accurately and faster by processing every pixel of the full image at once, rather than downscaling the image or sliding a window across it. To do this it introduces Dot localisation, a point-oriented detection method, and the Pixel Distill (PD) module, which halves resolution through pixel rearrangement while keeping all pixel information. It also introduces UP-COUNT, a new 4K drone dataset with 352,487 head annotations, a moving camera, and altitude metadata. On DroneCrowd the method reports state-of-the-art L-mAP 51.00 and L-AP@10 57.06; on UP-COUNT it reports 66.49 and 75.46, establishing a baseline for the new benchmark. If these numbers hold, the method offers a more accurate and a 3.85x faster alternative for drone-based crowd monitoring.

What carries the argument

The Pixel Distill (PD) module is the central mechanism: it applies a PixelUnshuffle layer to halve image resolution while increasing channels, splits the tensor into four parts along the channel axis, processes each with a block that includes Coordinate Attention, and concatenates the results before a channel-reduction convolution. This design lets a full 4K or Full HD frame be processed in one pass, with no interpolation downsampling and no sliding-window cropping, preserving spatial information of tiny objects. The other load-bearing piece is the PDL loss, which combines a modified Focal loss that creates Gaussian neighbourhoods around ground-truth points, a binary-entropy objectness term, and an L2 regression term measuring pixel distance to the nearest ground truth; the weights (0.25, 1, 2) were set empirically. Together these components produce masks from which object coordinates are extracted by 3x3 max-pooling NMS followed by thresholding.

What would settle it

A concrete check: train MFA on DroneCrowd under this paper's evaluation protocol and measure L-mAP; if it reaches or exceeds 51.00, the state-of-the-art claim is overturned.

Watch

Extended reading notes

Core claim

The central claim is that a point-oriented localisation pipeline built around the Pixel Distill module outperforms both prior drone-crowd-specific methods and adapted tiny-object detectors on two drone benchmarks. The authors argue that the main limitation of prior work is the way high-resolution images are handled: downscaling with interpolation discards information about tiny (2-8 pixel) people, while sliding-window inference loses global context and is computationally redundant. Dot localisation uses a UNet-style encoder-decoder with a Mix Vision Transformer backbone to produce a full-resolution mask, then extracts coordinates by non-maximum suppression and thresholding. Pixel Distill, inspired by PixelUnshuffle, rearranges pixels into multiple lower-resolution channel-split feature maps so that every pixel contributes to the representation, and the combined Point Distance-aware Localisation (PDL) loss balances modified focal loss, objectness, and Euclidean regression terms. With these components the method reports state-of-the-art results on DroneCrowd and strong performance on the new UP-COUNT dataset, while being more parameter-efficient and faster than the closest compared transformer detector.

Load-bearing premise

The state-of-the-art claim rests on the comparability of the quoted MFA numbers: the authors could not reproduce MFA on DroneCrowd and took its results from the original paper, so if those numbers came from a different evaluation protocol the claimed margin over the prior best method is not established.

Editorial extensions

If this is right

  • On DroneCrowd, Dot localisation with PD reports L-mAP 51.00 and L-AP@10 57.06, surpassing the previous state of the art by 7.57 and 9.92 points respectively.
  • On the new UP-COUNT dataset it establishes a baseline with L-mAP 66.49 and L-AP@10 75.46, and the dataset contributes 10,000 4K frames with 352,487 head annotations, moving-camera footage, and altitude information.
  • Ablation results show that replacing MSE with the PDL loss improves L-mAP by roughly 10.7 points averaged across datasets, and adding PD adds about 3.0 and 2.3 points on DroneCrowd and UP-COUNT respectively.
  • The full-image PD processing is more computationally efficient than sliding-window inference on 4K images; the complete model is 3.85x faster than the closest compared transformer detector, with 12 million fewer parameters.
  • Cross-dataset evaluation indicates that training on UP-COUNT transfers to DroneCrowd better than the reverse, and Dot outperforms the closest compared transformer detector in both transfer directions.

Reading between the lines

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

  • If the PD module is the reason for the gain, a similar pixel-rearrangement preprocessing could be applied to other dense small-object tasks, such as aerial vehicle counting or cell detection in microscopy, where full-resolution context matters but GPUs cannot fit the whole image.
  • The authors note that PD's resolution reduction is limited to factors of 2, 4, 8, and beyond; a testable extension is a learned or adaptive PixelUnshuffle factor that preserves more spatial detail for moderately high resolutions.
  • UP-COUNT's moving-camera and altitude metadata could enable future work that conditions detection on altitude or ego-motion, which the authors list as future work; the dataset makes such experiments possible for the first time.
  • The reported speedup suggests that full-image processing with pixel rearrangement may be a practical route to real-time drone-based crowd monitoring, but a direct end-to-end latency benchmark on the same hardware and software stack is needed to confirm it in deployment.
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

3 major / 5 minor

Summary. The paper proposes Dot localisation, a point-based object localisation method for high-resolution UAV imagery, together with a Pixel Distill (PD) module that processes full-resolution images without interpolation-based downscaling or sliding-window inference. The approach uses a UNet-like encoder-decoder with a MiT backbone, a composite loss (PDL) combining modified focal, objectness, and regression terms, and a post-processing step that extracts point coordinates from predicted masks. The authors also introduce UP-COUNT, a new 4K drone crowd localisation dataset with moving-camera footage, altitude information, and 352,487 head annotations. Experiments on DroneCrowd and UP-COUNT claim state-of-the-art results on DroneCrowd and establish a baseline on UP-COUNT, supplemented by ablation studies, cross-dataset generalisation tests, and a real-world use case.

Significance. If the reported results hold, the method offers a practical and efficient alternative to sliding-window and downscaling approaches for high-resolution drone imagery, with a claimed speed advantage and a public dataset that addresses realistic moving-camera conditions. The release of the UP-COUNT dataset and the availability of code are concrete contributions that will benefit the community. However, the central state-of-the-art claim currently rests on an un-reproduced baseline (MFA) and on headline numbers that are inconsistent between Table 2 and Table 3. These issues must be resolved before the empirical claims can be fully credited.

major comments (3)
  1. [Section 5, Tables 2 and 3] There is an internal numerical inconsistency for the same configuration on UP-COUNT. Table 2 reports Dot + PD with L-mAP 66.49 and L-AP@10 75.46, while Table 3 reports the MiT B2 + PDL + PD configuration with L-mAP 65.05 and L-AP@10 (labelled mAP@10) 74.03. Since the summary and conclusion cite the higher numbers to claim state-of-the-art results on UP-COUNT, this discrepancy is load-bearing. The authors must identify which numbers are correct, explain why the tables disagree, and ensure all final metrics are reported consistently throughout the paper.
  2. [Section 5, MFA baseline] The DroneCrowd state-of-the-art claim depends on the MFA results being quoted from Asanomi et al. (2023) because the authors state they could not reproduce MFA on DroneCrowd. The reported margin over MFA (7.57 L-mAP and 9.92 L-AP@10) is therefore not established under a common evaluation protocol. If MFA was evaluated with a different test split, matching rule, input resolution, or post-processing threshold, the comparison is not apples-to-apples. The authors should either reproduce MFA under the same protocol used for all other methods, or provide a detailed protocol comparison and explicitly qualify the state-of-the-art claim. The statement in Section 4.5 that 'We train the previous state-of-the-art methods using procedures and parameters provided by the authors' also needs to be reconciled with the statement that MFA could not be reproduced.
  3. [Section 5.3, Cross-dataset evaluation] The cross-dataset evaluation compares Dot with SD-DETR only. Given that the paper's main contribution is a new dataset, the generalisation analysis would be stronger if it also included the other baseline methods (e.g., STNNet, STEERER, RFLA) under the same transfer protocol. As it stands, the claim that 'Dot outperformed SD-DETR in both instances' is not sufficient to establish broader generalisation advantages over the other compared methods.
minor comments (5)
  1. [Table 3] The column heading 'mAP@10' should be 'L-AP@10' for consistency with Table 2 and the metric definitions in Section 5.
  2. [Section 5.1] The statement that 'the Dot model has 12M fewer parameters and provides 3.85x faster inference speed (0.013 vs 0.050 seconds on RTX4090)' does not specify the comparison method. Please state explicitly which architecture (presumably SD-DETR?) is used as the reference, and report the input resolution and batch size used for the timing measurement.
  3. [Section 4.4] The loss coefficients in Equation (1) are described as selected empirically, but no sensitivity analysis is provided. Since the ablation study attributes the improvement to PDL as a whole, it would be informative to report the effect of varying the coefficients, or at least state that the chosen values are fixed for all experiments.
  4. [Figure 3 caption] The caption says the PD module downsamples by two for DroneCrowd and by four for UP-COUNT, but Section 4.1 describes the second variant as having an additional PixelUnshuffle layer. Please align the terminology and clarify whether the downsampling factors refer to the overall module output resolution or to the internal processing.
  5. [Abstract and Section 1] The phrase 'utilising every pixel' is used repeatedly, but the Pixel Distill module actually applies PixelUnshuffle, channel splitting, and convolution, rather than treating each pixel independently in an explicit per-pixel computation. Consider rephrasing to 'processing the full-resolution image without information-discarding interpolation' to avoid overclaiming.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: core claims are empirical benchmark results on external datasets, with ablations, not derivations from fitted inputs.

full rationale

The paper is a systems/benchmark paper. The proposed Dot localisation and Pixel Distill module are evaluated on external datasets DroneCrowd and UP-COUNT; the reported performance numbers are measured, not derived from the method's own equations. The combined loss Ltotal = 0.25*Lneg + Lobj + 2*Lreg uses empirically selected coefficients, and the post-processing threshold (0.2) is a hyperparameter; neither is a fitted quantity renamed as a prediction. The only self-citation (Ptak et al. 2022) supports a general remark about downscaling information loss and is not load-bearing. The MFA baseline is quoted from Asanomi et al. (2023) because the authors could not reproduce it, which is a comparability/reproducibility caveat for the SOTA claim, not circularity, since the comparison target is external. One internal inconsistency is noted between Table 2 (UP-COUNT L-mAP 66.49 for Dot+PD) and Table 3 (same configuration 65.05), which is a correctness/consistency issue, not circularity. No load-bearing step reduces by construction to its input, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central empirical claims rest on standard deep-learning assumptions (pretraining, annotation quality) and on design choices such as loss weights and thresholds that are selected by hand. There are no invented physical entities; the main burden is on fair comparison with baselines.

free parameters (5)
  • Loss weights (0.25, 1, 2) = 0.25 for Lneg, 1 for Lobj, 2 for Lreg
    Section 4.4 states 'The constant coefficients were selected empirically.'
  • Mask threshold = 0.2
    Section 4.3: 'in our experiments, thresh = 0.2'.
  • PD downsampling factor = 2 for DroneCrowd, 4 for UP-COUNT
    Section 4.1 and Figure 3: chosen to match input resolution to UNet input; a design decision.
  • PD block channel count = 16 intermediate, 3 output channels
    Section 4.1: 'the sixteen-channel feature maps are extracted...' and 'This reduction is set to three channels'; chosen based on ImageNet pretraining benefits.
  • Learning rate, epochs, early stopping patience = 3e-4, 100 epochs, patience 20
    Section 4.5: standard hyperparameters.
assumptions (4)
  • domain assumption ImageNet pretraining improves downstream performance and is beneficial here
    Section 4.5: 'All evaluated architectures were pretrained on Imagenet'; also cited Huh et al. 2016 and Risojević and Stojnić 2021. The benefit is assumed, not measured in this paper.
  • domain assumption Head point annotations are reliable indicators of person location
    Section 3: labels were hand-prepared and checked by two people; the ground truth defines the evaluation, so errors would directly affect scores.
  • domain assumption PixelUnshuffle preserves all information needed for localization while reducing resolution
    Section 4.1: the module is built on this; the paper's ablation supports it empirically on two datasets, but it is an architectural assumption.
  • domain assumption Coordinate Attention improves feature extraction for this task
    Section 4.1: adapted from Hou et al. 2021; no independent verification in this paper beyond the full-model results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing people localisation in drone imagery for better crowd management by utilising every pixel in high-resolution images." pith.science (2026). https://pith.science/paper/BPYN5TX5

@misc{pith2026250204014,
  author       = {Pith},
  title        = {Pith review of: Enhancing people localisation in drone imagery for better crowd management by utilising every pixel in high-resolution images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BPYN5TX5}},
  note         = {Machine review of arXiv:2502.04014}
}
read the original abstract

Accurate people localisation using drones is crucial for effective crowd management, not only during massive events and public gatherings but also for monitoring daily urban crowd flow. Traditional methods for tiny object localisation using high-resolution drone imagery often face limitations in precision and efficiency, primarily due to constraints in image scaling and sliding window techniques. To address these challenges, a novel approach dedicated to point-oriented object localisation is proposed. Along with this approach, the Pixel Distill module is introduced to enhance the processing of high-definition images by extracting spatial information from individual pixels at once. Additionally, a new dataset named UP-COUNT, tailored to contemporary drone applications, is shared. It addresses a wide range of challenges in drone imagery, such as simultaneous camera and object movement during the image acquisition process, pushing forward the capabilities of crowd management applications. A comprehensive evaluation of the proposed method on the proposed dataset and the commonly used DroneCrowd dataset demonstrates the superiority of our approach over existing methods and highlights its efficacy in drone-based crowd object localisation tasks. These improvements markedly increase the algorithm's applicability to operate in real-world scenarios, enabling more reliable localisation and counting of individuals in dynamic environments.

Figures

Figures reproduced from arXiv: 2502.04014 by the authors.

Figure 1
Figure 1. A challenging example of tiny people localisation from UAV footage. While existing state-of-the-art methods [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Example images with drawn head labels from UP-COUNT dataset. Top - the lowest altitude; middle - the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Architecture of the designed Dot localisation approach. The Pixel Distill (PD) module processes a full [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visual comparison of results provided by different methods for both datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: L-AP metric comparison regarding the correctness threshold. The range of thresholds covers the L-mAP [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Four example frames from a video recording of a march. Each frame contains people detections marked in [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 47 canonical work pages

  1. [1]

    Applications of unmanned aerial vehicle ( UAV ) in road safety, traffic and highway infrastructure management: Recent advances and challenges

    Fatma Outay, Hanan Abdullah Mengash, and Muhammad Adnan. Applications of unmanned aerial vehicle ( UAV ) in road safety, traffic and highway infrastructure management: Recent advances and challenges. Transportation research part A: policy and practice, 141: 0 116--129, 2020

    work page 2020

  2. [2]

    Unmanned aerial vehicles as element of road traffic safety monitoring

    Olga Dronova, Denis Parinov, Bogdan Soloviev, Dinara Kasumova, Evgeniy Kochetkov, Olga Medvedeva, and Irina Sergeeva. Unmanned aerial vehicles as element of road traffic safety monitoring. Transportation research procedia, 63: 0 2308--2314, 2022

    work page 2022

  3. [3]

    Urban traffic monitoring and analysis using unmanned aerial vehicles ( UAVs ): A systematic literature review

    Eugen Valentin Butil a and R a zvan Gabriel Boboc. Urban traffic monitoring and analysis using unmanned aerial vehicles ( UAVs ): A systematic literature review. Remote Sensing, 14 0 (3): 0 620, 2022

    work page 2022

  4. [4]

    Unmanned aerial vehicles applications in future smart cities

    Nader Mohamed, Jameela Al-Jaroodi, Imad Jawhar, Ahmed Idries, and Farhan Mohammed. Unmanned aerial vehicles applications in future smart cities. Technological forecasting and social change, 153: 0 119293, 2020

    work page 2020

  5. [5]

    UAV fleet as a dependable service for smart cities: Model-based assessment and application

    Vyacheslav Kharchenko, Ihor Kliushnikov, Andrzej Rucinski, Herman Fesenko, and Oleg Illiashenko. UAV fleet as a dependable service for smart cities: Model-based assessment and application. Smart Cities, 5 0 (3): 0 1151--1178, 2022

    work page 2022

  6. [6]

    Object detection in optical remote sensing images: A survey and a new benchmark

    Ke Li, Gang Wan, Gong Cheng, Liqiu Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of Photogrammetry and Remote Sensing, 159: 0 296--307, 2020

    work page 2020

  7. [7]

    Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery

    Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184: 0 116--130, 2022

    work page 2022

  8. [8]

    Deep learning for small and tiny object detection: A survey

    Aleksandra Kos, Dominik Belter, and Karol Majek. Deep learning for small and tiny object detection: A survey. Pomiary Automatyka Robotyka, 27, 2023

    work page 2023

Show all 55 references
  1. [9]

    Unmanned aerial vehicles for crowd monitoring and analysis

    Muhammad Afif Husman, Waleed Albattah, Zulkifli Zainal Abidin, Yasir Mohd Mustafah, Kushsairy Kadir, Shabana Habib, Muhammad Islam, and Sheroz Khan. Unmanned aerial vehicles for crowd monitoring and analysis. Electronics, 10 0 (23): 0 2974, 2021

    2021

  2. [10]

    Unmanned aerial vehicle communications for civil applications: A review

    Mohammad Ghamari, Pablo Rangel, Mehrube Mehrubeoglu, Girma S Tewolde, and R Simon Sherratt. Unmanned aerial vehicle communications for civil applications: A review. IEEE Access, 10: 0 102492--102531, 2022

    2022

  3. [11]

    Assistance of uavs in the intelligent management of urban space: A survey

    Haoran Xu, Lizhe Wang, Wei Han, Yixin Yang, Jiabao Li, Yue Lu, and Jun Li. Assistance of uavs in the intelligent management of urban space: A survey. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023

    2023

  4. [12]

    UAV -based IoT platform: A crowd surveillance use case

    Naser Hossein Motlagh, Miloud Bagaa, and Tarik Taleb. UAV -based IoT platform: A crowd surveillance use case. IEEE Communications Magazine , 55 0 (2): 0 128--134, 2017

    2017

  5. [13]

    Development of automated people counting system using object detection and tracking

    Chee Jia Hong and Muhammad Hazli Mazlan. Development of automated people counting system using object detection and tracking. Inter. Journal of Online & Biomedical Engineering, 19 0 (6), 2023

    2023

  6. [14]

    A survey of object detection for uavs based on deep learning

    Guangyi Tang, Jianjun Ni, Yonghao Zhao, Yang Gu, and Weidong Cao. A survey of object detection for uavs based on deep learning. Remote Sensing, 16 0 (1): 0 149, 2023

    2023

  7. [15]

    Efficient high-resolution deep learning: A survey

    Arian Bakhtiarnia, Qi Zhang, and Alexandros Iosifidis. Efficient high-resolution deep learning: A survey. ACM Computing Surveys, 2022

    2022

  8. [16]

    On-board crowd counting and density estimation using low altitude unmanned aerial vehicles—looking beyond beating the benchmark

    Bartosz Ptak, Dominik Pieczy \'n ski, Mateusz Piechocki, and Marek Kraft. On-board crowd counting and density estimation using low altitude unmanned aerial vehicles—looking beyond beating the benchmark. Remote Sensing, 14 0 (10): 0 2288, 2022

    2022

  9. [17]

    An empirical study of context in object detection

    Santosh K Divvala, Derek Hoiem, James H Hays, Alexei A Efros, and Martial Hebert. An empirical study of context in object detection. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1271--1278. IEEE, 2009

    2009

  10. [18]

    Oriented ship detection based on soft thresholding and context information in SAR images of complex scenes

    Chuan Zhang, Gui Gao, Jia Liu, and Dingfeng Duan. Oriented ship detection based on soft thresholding and context information in SAR images of complex scenes. IEEE Trans. on Geoscience and Remote Sensing , 2023

    2023

  11. [19]

    Detection, tracking, and counting meets drones in crowds: A benchmark

    Longyin Wen, Dawei Du, Pengfei Zhu, Qinghua Hu, Qilong Wang, Liefeng Bo, and Siwei Lyu. Detection, tracking, and counting meets drones in crowds: A benchmark. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7812--7821, 2021

    2021

  12. [20]

    Context-aware crowd counting

    Weizhe Liu, Mathieu Salzmann, and Pascal Fua. Context-aware crowd counting. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5099--5108, 2019

    2019

  13. [21]

    CSRNet : Dilated convolutional neural networks for understanding the highly congested scenes

    Yuhong Li, Xiaofan Zhang, and Deming Chen. CSRNet : Dilated convolutional neural networks for understanding the highly congested scenes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1091--1100, 2018

    2018

  14. [22]

    Distribution matching for crowd counting

    Boyu Wang, Huidong Liu, Dimitris Samaras, and Minh Hoai Nguyen. Distribution matching for crowd counting. Advances in neural information processing systems, 33: 0 1595--1607, 2020

    2020

  15. [23]

    Rethinking counting and localization in crowds: A purely point-based framework

    Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3365--3374, 2021

    2021

  16. [24]

    An end-to-end transformer model for crowd localization

    Dingkang Liang, Wei Xu, and Xiang Bai. An end-to-end transformer model for crowd localization. In Eur. Conf. Comput. Vis., pages 38--54. Springer, 2022 a

    2022

  17. [25]

    Boosting detection in crowd analysis via underutilized output features

    Shaokai Wu and Fengyu Yang. Boosting detection in crowd analysis via underutilized output features. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15609--15618, 2023

    2023

  18. [26]

    STEERER : Resolving scale variations for counting and localization via selective inheritance learning

    Tao Han, Lei Bai, Lingbo Liu, and Wanli Ouyang. STEERER : Resolving scale variations for counting and localization via selective inheritance learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21848--21859, 2023

    2023

  19. [27]

    Multi-frame attention with feature-level warping for drone crowd tracking

    Takanori Asanomi, Kazuya Nishimura, and Ryoma Bise. Multi-frame attention with feature-level warping for drone crowd tracking. In Winter Conf. on Applications of Computer Vision, pages 1664--1673, 2023

    2023

  20. [28]

    U-net : Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net : Convolutional networks for biomedical image segmentation. In Inter. conf. on Medical image computing and computer-assisted intervention, pages 234--241. Springer, 2015

    2015

  21. [29]

    Tiny object detection in aerial images

    Jinwang Wang, Wen Yang, Haowen Guo, Ruixiang Zhang, and Gui-Song Xia. Tiny object detection in aerial images. In Int. Conf. Pattern Recog., pages 3791--3798. IEEE, 2021 a

    2021

  22. [30]

    A normalized gaussian wasserstein distance for tiny object detection

    J Wang, C Xu, W Yang, and L Yu. A normalized gaussian wasserstein distance for tiny object detection. arXiv preprint arXiv:2110.13389, 2021 b

    2021 arXiv

  23. [31]

    RFLA : Gaussian receptive field based label assignment for tiny object detection

    Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. RFLA : Gaussian receptive field based label assignment for tiny object detection. In Eur. Conf. Comput. Vis., pages 526--543. Springer, 2022

    2022

  24. [32]

    A transformer-based framework for tiny object detection

    Yi-Kai Liao, Gong-Si Lin, and Mei-Chen Yeh. A transformer-based framework for tiny object detection. In Asia Pacific Signal and Inf. Proc. Association Annual Summit and Conf. (APSIPA ASC), pages 373--377. IEEE, 2023

    2023

  25. [33]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll \'a r. Focal loss for dense object detection. In Int. Conf. Comput. Vis., pages 2980--2988, 2017

    2017

  26. [34]

    Focal inverse distance transform maps for crowd localization

    Dingkang Liang, Wei Xu, Yingying Zhu, and Yu Zhou. Focal inverse distance transform maps for crowd localization. IEEE Trans. on Multimedia, 2022 b

    2022

  27. [35]

    Mask focal loss for dense crowd counting with canonical object detection networks

    Xiaopin Zhong, Guankun Wang, Weixiang Liua, Zongze Wua, and Yuanlong Deng. Mask focal loss for dense crowd counting with canonical object detection networks. arXiv preprint arXiv:2212.11542, 2022

    2022 arXiv

  28. [36]

    Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network

    Wenzhe Shi, Jose Caballero, Ferenc Husz \'a r, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In IEEE Conf. Comput. Vis. Pattern Recog., p...

    2016

  29. [37]

    Coordinate attention for efficient mobile network design

    Qibin Hou, Daquan Zhou, and Jiashi Feng. Coordinate attention for efficient mobile network design. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13713--13722, 2021

    2021

  30. [38]

    What makes imagenet good for transfer learning? arXiv preprint arXiv:1608.08614, 2016

    Minyoung Huh, Pulkit Agrawal, and Alexei A Efros. What makes imagenet good for transfer learning? arXiv preprint arXiv:1608.08614, 2016

    2016 arXiv

  31. [39]

    Do we still need imagenet pre-training in remote sensing scene classification? arXiv preprint arXiv:2111.03690, 2021

    Vladimir Risojevi \'c and Vladan Stojni \'c . Do we still need imagenet pre-training in remote sensing scene classification? arXiv preprint arXiv:2111.03690, 2021

    2021 arXiv

  32. [40]

    Swin transformer embedding unet for remote sensing image semantic segmentation

    Xin He, Yong Zhou, Jiaqi Zhao, Di Zhang, Rui Yao, and Yong Xue. Swin transformer embedding unet for remote sensing image semantic segmentation. IEEE Trans. on Geoscience and Remote Sensing, 60: 0 1--15, 2022

    2022

  33. [41]

    SegFormer : Simple and efficient design for semantic segmentation with transformers

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. SegFormer : Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34: 0 12077--12090, 2021

    2021

  34. [42]

    Cornernet: Detecting objects as paired keypoints

    Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proc. of the European conf. on computer vision (ECCV), pages 734--750, 2018

    2018

  35. [43]

    YOLOv7 : Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors

    Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. YOLOv7 : Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7464--7475, 2023

    2023

  36. [44]

    YOLOv4 : Optimal speed and accuracy of object detection

    Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. YOLOv4 : Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020

    2004 arXiv

  37. [45]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  38. [46]

    Attention mechanisms in computer vision: A survey

    Meng-Hao Guo, Tian-Xing Xu, Jiang-Jiang Liu, Zheng-Ning Liu, Peng-Tao Jiang, Tai-Jiang Mu, Song-Hai Zhang, Ralph R Martin, Ming-Ming Cheng, and Shi-Min Hu. Attention mechanisms in computer vision: A survey. Computational Visual Media, 8 0 (3): 0 331--368, 2022

    2022

  39. [47]

    SGDR : Stochastic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. SGDR : Stochastic gradient descent with warm restarts. arXiv arXiv:1608.03983, 2016

    2016 arXiv

  40. [48]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., pages 248--255. IEEE, 2009

    2009

  41. [49]

    Training deeper convolutional networks with deep supervision

    Liwei Wang, Chen-Yu Lee, Zhuowen Tu, and Svetlana Lazebnik. Training deeper convolutional networks with deep supervision. arXiv preprint arXiv:1505.02496, 2015

    2015 arXiv

  42. [50]

    Object detection in aerial images: A large-scale benchmark and challenges

    Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Object detection in aerial images: A large-scale benchmark and challenges. IEEE Trans. on Pattern Analysis and Machine Intelligenc...

    2021

  43. [51]

    Towards large-scale small object detection: Survey and benchmarks

    Gong Cheng, Xiang Yuan, Xiwen Yao, Kebing Yan, Qinghua Zeng, Xingxing Xie, and Junwei Han. Towards large-scale small object detection: Survey and benchmarks. IEEE Trans. on Pattern Analysis and Machine Intelligence, 2023

    2023

  44. [52]

    The unmanned aerial vehicle benchmark: Object detection and tracking

    Dawei Du, Yuankai Qi, Hongyang Yu, Yifan Yang, Kaiwen Duan, Guorong Li, Weigang Zhang, Qingming Huang, and Qi Tian. The unmanned aerial vehicle benchmark: Object detection and tracking. In Eur. Conf. Comput. Vis., pages 370--386, 2018

    2018

  45. [53]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770--778, 2016

    2016

  46. [54]

    Resnet strikes back: An improved training procedure in timm

    Ross Wightman, Hugo Touvron, and Herv \'e J \'e gou. Resnet strikes back: An improved training procedure in timm. arXiv preprint arXiv:2110.00476, 2021

    2021 arXiv

  47. [55]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Inter. Conf. on Machine Learning, pages 6105--6114. PMLR, 2019

    2019

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.