Pith. sign in

REVIEW 3 major objections 6 minor 32 references

YOLOv8-SMOT: An Efficient and Robust Framework for Real-Time Small Object Tracking via Slice-Assisted Training and Adaptive Association

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read An appearance-free tracker for small moving objects, trained on sliced high-resolution frames with a YOLOv8 detector, reports state-of-the-art SO-HOTA of 55.205 on the SMOT4SB public test set.

desk verdict Challenge-winning tracker with a real headline number, but SliceTrain's own contribution is never isolated in an ablation, and the public-test comparison confounds detector scale, resolution, and training strategy. read the letter →

arxiv 2507.12087 v2 pith:FV33JI6P submitted 2025-07-16 cs.CV

Xiang Yu , Xinyao Liu , Guang Liang This is my paper
classification cs.CV
keywords smallobjecttrackingmulti-objectUAVaerialYOLOv8SliceTrainOC-SORTdataassociationmotionprediction
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

Tracking small, agile birds from a moving drone is hard because targets have almost no usable appearance, the camera and target motions are entangled, and flocks cause frequent occlusions. This paper argues that a tracking-by-detection system can handle all three without any appearance model: train the YOLOv8 detector on overlapping, fully covering slices of high-resolution frames with slice-level augmentation, and associate detections with a motion-only tracker built on OC-SORT that keeps an EMA of moving direction and scores matches by expanded-IoU minus a normalized center distance. On the SMOT4SB public test set the full system with the YOLOv8-L detector reports SO-HOTA 55.205, up from 44.200 for the default OC-SORT tracker with the same detector. The gain matters because it shows a practical path to real-time small-object tracking where classic re-identification strategies are known to fail.

What carries the argument

Two mechanisms carry the argument. SliceTrain is a training-time data pipeline: it tiles each high-resolution frame (e.g., 2160×3840) into overlapping 1280×1280 tiles so every pixel is covered, applies stochastic augmentation independently per tile, and thereby raises the usable batch size while keeping information density high; inference then runs on the original full image, avoiding stitching. The tracker is an appearance-free extension of OC-SORT whose matching cost is an adaptive similarity, $$\text{Similarity} = \frac{\text{ExpandedIoU} - \text{NormalizedDistance} + 1}{2},$$ with bounding boxes expanded by a factor of 2 before IoU, and whose motion prediction uses an EMA velocity, $v^t_{EMA} = \alpha v^{t-1}_{EMA} + (1-\alpha)v^{t-1}_{ins}$, to smooth over sudden turns.

What would settle it

Run the same system on a held-out video set in which a majority of bird instances change speed by more than 20% between consecutive frames or jump more than twice their box width/height; if the reported gains from EMA smoothing and the distance penalty vanish or reverse, then the robust-motion claim does not extend to erratic-motion regimes.

Watch

Extended reading notes

Core claim

The paper's central claim is that the bottleneck in small multi-object tracking is not detection resolution alone, nor association alone, but the interaction between them—and that both sides can be fixed with cheap, modular changes. On the detection side, jointly applying deterministic full-coverage tiling and slice-level stochastic augmentation lets a standard YOLOv8 detector learn from high-resolution imagery at practical batch sizes, then run full-size inference without stitching. On the association side, the system shows that replacing raw IoU with a more reliable similarity—expanded bounding boxes plus a center-distance penalty, guided by an EMA-smoothed motion direction—supplies the missing signal exactly in the regime where raw IoU is near zero. The reported results, with each component added in sequence raising SO-HOTA from 44.200 to 55.205 on the public test set, are offered as evidence that the combination transfers to a real drone-captured flocking-bird benchmark.

Load-bearing premise

The tracker's success rests on assuming that the motion statistics measured on the training videos—mostly near-constant speed and per-frame displacement within roughly twice the box size—also hold on the test videos.

Editorial extensions

If this is right

  • Small-object detectors can be trained on sliced high-resolution data at large batch sizes and still be applied to full images, so high-resolution aerial detection no longer requires huge GPU memory at training time.
  • A tracker that never uses appearance can hold identities through dense flocks and brief occlusions, provided the motion signal is made reliable with EMA smoothing and an expanded matching space.
  • Because the distance penalty contributed the largest single jump in SO-HOTA, further improving the similarity metric is a promising lever for small-object tracking without re-identification.
  • The smaller YOLOv8-S variant retains most of the accuracy at roughly three times the speed, so the framework can be tuned toward real-time deployment on drones.

Reading between the lines

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

  • A natural extension of SliceTrain is to other small-object detection tasks in aerial or satellite imagery; a direct test would be applying it to existing small-object benchmarks with similar target-to-tile scale.
  • The similarity metric is modular and could be plugged into other observation-centric trackers; comparing it against plain IoU on identical detections would isolate its contribution from EMA's.
  • Because the motion assumptions are frame-rate dependent, a sensible stress test is low-frame-rate drone footage, where near-constant velocity between consecutive frames is more likely to fail.
  • The paper's motion-statistics tables also work as a cheap pre-deployment diagnostic: measure velocity-ratio and displacement distributions before trusting an appearance-free tracker in a new setting.
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 / 6 minor

Summary. The paper presents YOLOv8-SMOT, a tracking-by-detection framework for the SMOT4SB bird-tracking challenge. The detection side uses SliceTrain, a two-step training pipeline that tiles high-resolution images into overlapping patches and applies slice-level stochastic augmentation before fine-tuning YOLOv8 at several scales, with inference performed on full-size frames. The tracking side builds on OC-SORT and adds an EMA-smoothed motion direction plus an adaptive similarity metric that combines bounding-box expansion with a normalized center-distance penalty. On the public SMOT4SB test set the authors report SO-HOTA 55.205 for YOLOv8-L, with a monotonic ablation sequence from 44.200 to 55.205 in Table 4.

Significance. If the reported results are reproducible, the work is a useful engineering contribution: it shows that an appearance-free tracker with simple motion cues can reach strong performance in a challenging small-object tracking benchmark, and the per-component tracker diagnostics in Table 3 are a nice way to justify the similarity metric. The use of the official SO-HOTA metric and public test set makes the headline numbers meaningful. The main weakness is that the detection-side contribution, SliceTrain, is not supported by any controlled experiment; the only comparison confounds detector scale, training procedure, inference resolution, and tracker. Code is promised but not yet released, and all ablation numbers are single-run, so the current evidence is suggestive rather than conclusive.

major comments (3)
  1. [3.1, Table 2] SliceTrain is never isolated in the experiments. Table 2 compares three YOLOv8 scales trained with SliceTrain against an external baseline that was evaluated at 1792x3264 instead of the 2160x3840 used for the proposed models, so the comparison confounds detector scale, training pipeline, inference resolution, and the tracker. Because Table 4's ablations hold the SliceTrain-trained detector fixed, the reported SO-HOTA 55.205 does not establish that SliceTrain contributes anything to the final result. Please add an experiment that trains the same detector architecture with and without SliceTrain (and ideally with a random-crop or SAHI fine-tuning control) at the same inference resolution and with the same tracker.
  2. [4.4, Table 4] The tracker ablation is presented as a single evaluation on the public test set with no variance estimates. The incremental gains in SO-HOTA (3.716 for +EMA, 3.471 for +BBox Expansion, 3.818 for +Dist Penalty) are plausible, but since SliceTrain's stochastic augmentations and the detector training are random, one cannot rule out that these differences are within run-to-run noise. Please report results over multiple training seeds or at least a validation-set mean and standard deviation.
  3. [3.2.1, 3.2.3, Figure 4] The tracker's hyperparameters (alpha=0.8, expansion scale=2, IoU thresholds) are selected from training-set motion statistics. The paper states that 76.2% of training instances satisfy the near-constant-velocity assumption and that most displacements fall within twice the box size, but it offers no evidence that these statistics transfer to the test distribution. Given the paper's own admission in Section 5 that the method relies on motion heuristics, please add a discussion of how sensitive the results are to these assumptions, or a robustness check on a subset of validation videos with more erratic motion.
minor comments (6)
  1. [3.2.2, Eq. (3.1)] The EMA update uses v_{t-1}^{ins} rather than v_t^{ins}; please clarify whether this one-frame lag is intentional to avoid using the current observation.
  2. [4.4, Table 4] Please state explicitly what the 'Default' row is (presumably OC-SORT with the same SliceTrain-trained YOLOv8-L detector) and list any post-processing steps shared across all rows.
  3. [Figure 4] The caption says 'horizontal (left) or vertical (right) displacement to width (left) or height (right)', but the axes, units, and the proportion of instances covered are not described; adding axis labels and a note on the covered fraction would make the figure self-contained.
  4. [Abstract, Table 2] The abstract and Section 1 call the system 'real-time', while Table 2 reports 5.70 FPS for the best-scoring YOLOv8-L model; the S variant reaches 17.61 FPS, so 'near real-time' would be more accurate.
  5. [3.2.3, Eq. (3.2)] Eq. (3.2) does not define NormalizedDistance; please state explicitly that it is the L2 center distance normalized by some reference length (e.g., the expanded-box diagonal or the box size).
  6. [3.2.3, Reference [24]] Reference [24] is closely related to the bounding-box expansion idea; please explain in Section 3.2.3 how the proposed expansion differs from the buffered matching space of [24].

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all reported gains are empirical evaluations on an external public test set, and the tuned hyperparameters are fitted to training data, not renamed as predictions.

full rationale

The paper makes no first-principles derivation; its claims are empirical. The tracker hyperparameters (IoU threshold 0.25, decrement 0.08, alpha=0.8, expansion scale 2) are explicitly tuned on the SMOT4SB training set by grid search and cross-validation, and then evaluated on the held-out public test set in Tables 2 and 4. The design of the adaptive similarity metric is motivated by training-set statistics (Table 3, Figure 4), but this is a standard fit-then-test pipeline, not a prediction that reduces to its own inputs by construction. The only self-citations are to the challenge summary paper [9], which defines the dataset and metric and includes the authors among many co-authors, and to GPLQ [11], which is mentioned only as a possible future quantization tool and is not load-bearing. The lack of an ablation isolating SliceTrain from detector scale and inference resolution is an experimental-support weakness, not circularity. The conclusion explicitly acknowledges reliance on motion heuristics as a limitation. Under the rule that a non-finding is expected for externally benchmarked engineering papers, the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central empirical claim rests on a small set of hyperparameters tuned on the training set plus two statistical assumptions about bird motion. No new physical or conceptual entities are introduced; SliceTrain, EMA, and the similarity metric are methods built from existing components. The external validity of the score depends on the test benchmark and on the transfer of training-set motion statistics.

free parameters (6)
  • Slice tile size and overlap ratio = 1280 x 1280, overlap 20%
    Chosen to fit GPU memory and increase batch size from 1 to 6 (Section 4.2); no sensitivity analysis reported.
  • IoU matching threshold = 0.25
    Optimal value determined by grid search on the training set (Section 4.2).
  • IoU threshold decrement for stages 2 and 3 = 0.08
    Optimal value from the same grid search (Section 4.2).
  • Detection confidence track threshold = 0.25
    Set based on detector outputs on the training set (Section 4.2).
  • EMA alpha = 0.8
    Chosen by cross-validation on the training set (Section 4.2).
  • Bounding box expansion scale = 2
    Chosen from empirical observation that most inter-frame displacements do not exceed twice the box width or height (Section 4.2, Figure 4).
assumptions (5)
  • standard math The exponential moving average recursion in Eq. 3.1 is well-defined and stable for alpha in [0,1].
    Standard smoothing identity; no independent evidence needed.
  • domain assumption Most bird motion is locally near-constant-velocity, so a linear-motion tracker is appropriate.
    Supported only by training-set statistics in Table 1 (76.2% within one-frame window, 64.2% within four-frame window); assumed to hold on test data.
  • domain assumption A fixed expansion scale of 2 for bounding boxes is enough to make small-object matches overlap.
    Chosen from training-set displacement distributions in Figure 4; no sensitivity analysis and no guarantee on test distribution.
  • domain assumption The SO-HOTA metric with Dot Distance is a valid evaluation of tracking quality and the reported public test scores are accurate.
    The paper relies on the challenge's official metric and scores; it does not provide code to verify these numbers.
  • standard math Standard IoU and DIoU definitions are used as referenced in Eq. 3.2 and Table 3.
    Standard definitions from prior work [29]; no new content.

how reviews work

0 comments
Cite this review

Pith. "Pith review of YOLOv8-SMOT: An Efficient and Robust Framework for Real-Time Small Object Tracking via Slice-Assisted Training and Adaptive Association." pith.science (2026). https://pith.science/paper/FV33JI6P

@misc{pith2026250712087,
  author       = {Pith},
  title        = {Pith review of: YOLOv8-SMOT: An Efficient and Robust Framework for Real-Time Small Object Tracking via Slice-Assisted Training and Adaptive Association},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FV33JI6P}},
  note         = {Machine review of arXiv:2507.12087}
}
read the original abstract

Tracking small, agile multi-objects (SMOT), such as birds, from an Unmanned Aerial Vehicle (UAV) perspective is a highly challenging computer vision task. The difficulty stems from three main sources: the extreme scarcity of target appearance features, the complex motion entanglement caused by the combined dynamics of the camera and the targets themselves, and the frequent occlusions and identity ambiguity arising from dense flocking behavior. This paper details our championship-winning solution in the MVA 2025 "Finding Birds" Small Multi-Object Tracking Challenge (SMOT4SB), which adopts the tracking-by-detection paradigm with targeted innovations at both the detection and association levels. On the detection side, we propose a systematic training enhancement framework named \textbf{SliceTrain}. This framework, through the synergy of 'deterministic full-coverage slicing' and 'slice-level stochastic augmentation, effectively addresses the problem of insufficient learning for small objects in high-resolution image training. On the tracking side, we designed a robust tracker that is completely independent of appearance information. By integrating a \textbf{motion direction maintenance (EMA)} mechanism and an \textbf{adaptive similarity metric} combining \textbf{bounding box expansion and distance penalty} into the OC-SORT framework, our tracker can stably handle irregular motion and maintain target identities. Our method achieves state-of-the-art performance on the SMOT4SB public test set, reaching an SO-HOTA score of \textbf{55.205}, which fully validates the effectiveness and advancement of our framework in solving complex real-world SMOT problems. The source code will be made available at https://github.com/Salvatore-Love/YOLOv8-SMOT.

Figures

Figures reproduced from arXiv: 2507.12087 by the authors.

Figure 1
Figure 1. Model Overview. 3.1.1 SliceTrain Framework: Overcoming the Resolution-Diversity Dilemma High-quality small object detection, when dealing with high-resolution images, generally faces a core predicament: The Resolution-Diversity Dilemma. On one hand, high-resolution input (e.g., 3840 × 2160) is key to capturing the details of tiny objects. On the other hand, the enormous memory overhead limits the batch size to extre… view at source ↗
Figure 2
Figure 2. Distribution of velocity of all instances of training dataset. The instance ratio is [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Example of IoU caculation of object A and object B after expanding bounding boxes. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The ratio of horizontal (left) or vertical (right) displacement to width (left) or height [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 20 canonical work pages

  1. [24]

    Hard to Track Objects with Irregular Motions and Similar Appearances? Make It Easier by Buffering the Matching Space

    Fan Yang, Shigeyuki Odashima, Shoichi Masui, and Shan Jiang. Hard to Track Objects with Irregular Motions and Similar Appearances? Make It Easier by Buffering the Matching Space, November 2023. 83 citations (Semantic Scholar/DOI) [2025-04-23] arXiv:2211.14317 [cs]

    work page Pith review arXiv 2023

  2. [29]

    Distance- IoU Loss: Faster and Better Learning for Bounding Box Regression, November 2019

    Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance- IoU Loss: Faster and Better Learning for Bounding Box Regression, November 2019. arXiv:1911.08287 [cs]

    arXiv 2019

  3. [1]

    Slicing aided hyper inference and fine-tuning for small object detection

    Fatih Cagatay Akyon, Sinan Onur Altinuc, and Alptekin Temizel. Slicing aided hyper inference and fine-tuning for small object detection. In 2022 IEEE international conference on image processing (ICIP), pages 966–970. IEEE, 2022

    work page 2022

  4. [2]

    Simple online and realtime tracking

    Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464–3468. Ieee, 2016

    2016

  5. [3]

    Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking, March

    Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking, March

    work page

  6. [4]

    MOT20: A benchmark for multi object tracking in crowded scenes, March 2020

    Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taix´ e. MOT20: A benchmark for multi object tracking in crowded scenes, March 2020. arXiv:2003.09003 [cs]

    arXiv 2020

  7. [5]

    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 Proceedings of the European conference on computer vision (ECCV), pages 370–386, 2018

    work page 2018

  8. [6]

    Quantization without tears

    Minghao Fu, Hao Yu, Jie Shao, Junjie Zhou, Ke Zhu, and Jianxin Wu. Quantization without tears. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 4462–4472, 2025

    work page 2025

Show all 32 references
  1. [7]

    Ultralytics YOLO, January 2023

    Glenn Jocher, Jing Qiu, and Ayush Chaurasia. Ultralytics YOLO, January 2023

    2023

  2. [8]

    Baseline code for SMOT4SB by IIM-TTIJ, 2025

    Riku Kanayama, Yuki Yoshida, and Yuki Kondo. Baseline code for SMOT4SB by IIM-TTIJ, 2025

    2025

  3. [9]

    MV A 2025 Small Multi-Object Tracking for Spotting Birds Challenge: Dataset, Methods, and Results

    Yuki Kondo, Norimichi Ukita, Riku Kanayama, Yuki Yoshida, Takayuki Yamaguchi, Xi- ang Yu, Guang Liang, Xinyao Liu, Guan-Zhang Wang, Wei-Ta Chu, Bing-Cheng Chuang, Jia-Hua Lee, Pin-Tseng Kuo, I-Hsuan Chu, Yi-Shein Hsiao, Cheng-Han Wu, Po-Yi Wu, Jui-Chien Tsou, Hsuan-Chi Liu, Ch...

    2025

  4. [10]

    Mva2023 small object detection challenge for spotting birds: Dataset, methods, and results

    Yuki Kondo, Norimichi Ukita, Takayuki Yamaguchi, Hao-Yu Hou, Mu-Yi Shen, Chia-Chi Hsu, En-Ming Huang, Yu-Chen Huang, Yu-Cheng Xia, Chien-Yao Wang, et al. Mva2023 small object detection challenge for spotting birds: Dataset, methods, and results. In 2023 18th International Conf...

    2023

  5. [11]

    Gplq: A general, practical, and lightning qat method for vision transformers

    Guang Liang, Xinyao Liu, and Jianxin Wu. Gplq: A general, practical, and lightning qat method for vision transformers. arXiv preprint arXiv:2506.11784, 2025

    2025 arXiv

  6. [12]

    Feature pyramid networks for object detection

    Tsung-Yi Lin, Piotr Doll´ ar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117–2125, 2017. 10

    2017

  7. [13]

    A survey and performance evaluation of deep learning methods for small object detection

    Yang Liu, Peng Sun, Nickolas Wergeles, and Yi Shang. A survey and performance evaluation of deep learning methods for small object detection. Expert Systems with Applications, 172:114602, 2021

    2021

  8. [14]

    HOTA: A Higher Order Metric for Evaluating Multi- Object Tracking

    Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. HOTA: A Higher Order Metric for Evaluating Multi- Object Tracking. International Journal of Computer Vision, 129(2):548–578, February

  9. [16]

    MOT16: A Benchmark for Multi-Object Tracking, May 2016

    Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. MOT16: A Benchmark for Multi-Object Tracking, May 2016. arXiv:1603.00831 [cs]

    2016 arXiv

  10. [17]

    The power of tiling for small object detection

    F Ozge Unel, Burak O Ozkalayci, and Cevahir Cigla. The power of tiling for small object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019

    2019

  11. [18]

    Flocks, herds and schools: A distributed behavioral model

    Craig W Reynolds. Flocks, herds and schools: A distributed behavioral model. In Proceed- ings of the 14th annual conference on Computer graphics and interactive techniques, pages 25–34, 1987

    1987

  12. [19]

    Transtrack: Multiple object tracking with transformer

    Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020

    2012 arXiv

  13. [20]

    Qwt-v2: Practical, effective and efficient post-training quantization

    Ningyuan Tang, Minghao Fu, Hao Yu, and Jianxin Wu. Qwt-v2: Practical, effective and efficient post-training quantization. arXiv preprint arXiv:2505.20932, 2025

    2025 arXiv

  14. [21]

    Towards real-time multi-object tracking

    Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, and Shengjin Wang. Towards real-time multi-object tracking. In European conference on computer vision, pages 107–122. Springer, 2020

    2020

  15. [22]

    Simple online and realtime tracking with a deep association metric

    Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP), pages 3645–3649. IEEE, 2017

    2017

  16. [23]

    Dot Distance for Tiny Object Detec- tion in Aerial Images

    Chang Xu, Jinwang Wang, Wen Yang, and Lei Yu. Dot Distance for Tiny Object Detec- tion in Aerial Images. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPR W), pages 1192–1201, Nashville, TN, USA, June 2021. IEEE

    2021

  17. [25]

    A unified object motion and affinity model for online multi-object tracking

    Junbo Yin, Wenguan Wang, Qinghao Meng, Ruigang Yang, and Jianbing Shen. A unified object motion and affinity model for online multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6768–6777, 2020

    2020

  18. [26]

    Motr: End-to-end multiple-object tracking with transformer

    Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In European conference on computer vision, pages 659–675. Springer, 2022. 11

    2022

  19. [27]

    ByteTrack: Multi-Object Tracking by Associating Every Detection Box, April 2022

    Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. ByteTrack: Multi-Object Tracking by Associating Every Detection Box, April 2022. arXiv:2110.06864 [cs]

    2022 arXiv

  20. [28]

    Fairmot: On the fairness of detection and re-identification in multiple object tracking

    Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International journal of computer vision, 129(11):3069–3087, 2021

    2021

  21. [30]

    Vision meets drones: A challenge

    Pengfei Zhu, Longyin Wen, Xiao Bian, Haibin Ling, and Qinghua Hu. Vision meets drones: A challenge. arXiv preprint arXiv:1804.07437, 2018

    2018 arXiv

  22. [31]

    Learning data augmentation strategies for object detection

    Barret Zoph, Ekin D Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V Le. Learning data augmentation strategies for object detection. In European conference on computer vision, pages 566–583. Springer, 2020. 12

    2020

  23. [2021]

    arXiv:2009.07736 [cs] TLDR: This work presents a novel MOT evaluation metric, higher order tracking accuracy (HOTA), which explicitly balances the effect of performing accurate detection, association and localization into a single unified metric for comparing trackers

    2009 arXiv

  24. [2023]

    arXiv:2203.14360 [cs]

    arXiv

Pith tools

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