Pith. sign in

REVIEW 5 major objections 8 minor 23 references

Context-Aware Outlier Rejection for Robust Multi-View 3D Tracking of Similar Small Birds in An Outdoor Aviary

T0 review · 5 major / 8 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read A Voronoi landmark filter hits 97% match accuracy for look-alike birds in multi-view 3D tracking.

desk verdict Sensible Voronoi-based outlier rejection idea and a useful multi-bird dataset, but the evaluation is circular and the tracking claims are unsupported. read the letter →

arxiv 2412.16511 v1 pith:EHGPRCSI submitted 2024-12-21 cs.CV

classification cs.CV
keywords multi-viewtrackingoutlierrejectionVoronoidiagramfeaturematching3Dreconstructionbirdbehaviorhousesparrowaviarydataset
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 establish that a multi-view 3D tracking pipeline can reliably track many visually identical small birds in an outdoor aviary if it uses the environment, not just appearance, to decide which feature matches are real. The authors propose a context-aware outlier rejection step: a Voronoi diagram built from fixed environmental landmarks is layered over each camera view, and a feature match between two views is kept only when the matched keypoints fall in the same landmark cell on both sides. In experiments on 20 hours of footage of 80 house sparrows, the method rejected about 79% of initial brute-force matches and raised the ratio of correct matches among the survivors to 97%, yielding 3D reconstructions with a mean reprojection error below 20 pixels and more consistent tracks. If the claim holds, it offers a practical way to study social behavior and movement of wild birds in near-natural conditions without tagging individuals.

What carries the argument

The load-bearing object is the Voronoi diagram of environmental landmarks. In each camera view, a set of fixed landmarks is selected, and the image is tessellated into cells whose points share a common nearest landmark under Euclidean distance; virtual landmarks are added outside the image so no cell extends to infinity. The outlier rejection rule is then: a keypoint correspondence between two views is kept only if the two keypoints belong to Voronoi cells of the same landmark, so the nearest landmark agrees across views. This turns the problem of telling similar birds apart into a check of spatial context rather than appearance.

What would settle it

Compare keypoint correspondences against ground-truth matches in frames where a bird sits near a Voronoi cell boundary; if a large fraction of true matches is rejected because perspective makes the two views disagree about the nearest landmark, then the filter trades away real correspondences and its 97% precision is inflated by the high rejection rate.

Watch

Extended reading notes

Core claim

The central discovery is that environmental geometry—specifically the Voronoi partition of each camera image induced by a handful of user-chosen landmarks—carries enough information to separate correct correspondences from the large number of false matches produced by appearance-based matching of similar birds. For a match to be valid, the nearest landmark to the keypoint in one camera must be the same as the nearest landmark to its matched keypoint in the other camera; otherwise the match is rejected. This simple cell-agreement rule, combined with SIFT keypoints and brute-force k-NN matching, drives the reported 97% matching accuracy, a roughly 79% outlier rejection rate, and the subsequent triangulation-based 3D tracking with Kalman smoothing.

Load-bearing premise

The method assumes that a keypoint's nearest landmark in one camera view is also its nearest landmark in the other camera view, so a disagreement can be taken as proof the match is wrong.

Editorial extensions

If this is right

  • Multi-view tracking of visually similar birds becomes feasible without per-bird appearance models, since spatial context does the disambiguation.
  • The outlier rejection step removes roughly 79% of initial feature matches, and the surviving matches are about 97% correct, which is what allows accurate 3D reconstruction.
  • The pipeline tracks birds consistently for tens of seconds with an average of about 23 ID switches per minute, and over 77% of birds remain tracked for at least 10 seconds.
  • The released dataset of 80 banded house sparrows in four enclosures, with 20 hours of five-view footage, provides a testbed for other tracking and behavior-analysis methods.

Reading between the lines

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

  • The same nearest-landmark agreement rule could generalize to other multi-view tracking problems where objects look alike but move in a fixed scene, such as fish in a tank or rodents in an arena, as long as enough stable landmarks are visible.
  • Because landmarks are user-chosen, the method's robustness likely depends on landmark placement and density; placing landmarks too sparsely or too close to frequently used perches could change the rejection rate and matching accuracy, a sensitivity the paper does not quantify.
  • If the rejection filter is as effective as reported, it could also be used as a pre-filter inside learning-based matchers, reducing the burden on learned descriptors rather than replacing them.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 8 minor

Summary. The paper presents a multi-view 3D multi-bird tracking pipeline for an outdoor aviary that combines YOLOv5-based detection, SIFT keypoint extraction, brute-force feature matching, and a Voronoi-diagram-based outlier rejection step that uses user-defined environmental landmarks as context. The authors report a 97% matching accuracy, a rejection rate of roughly 20% in the abstract but 79% in the experiments, and tracking metrics including reprojection errors and ID-switch counts. They also release a dataset of 80 house sparrows over 20 hours of footage. The central claims are that the landmark-based context improves differentiation of visually similar birds and yields robust and reliable multi-bird 3D tracking.

Significance. The application domain is timely and the dataset release is a potentially valuable asset for computer vision, ornithology, and ecology. If the central claims were supported, the Voronoi-landmark outlier rejection idea would be a simple and interesting mechanism for exploiting static environmental structure in multi-view animal tracking. However, the experimental evaluation as written does not establish these claims: the matching accuracy is defined self-referentially without independent ground truth, no baseline comparison is provided, and the reported tracking retention contradicts the 'robust and reliable' conclusion. The paper's contribution is therefore currently more of a system description with an unvalidated central metric than a demonstrated method.

major comments (5)
  1. [§9.1, Table 3, Abstract] The reported '97% accuracy in matching' is not supported as an accuracy measure. Table 3 defines 'Ratio correct final matches / all final matches' without any external ground truth; a final match is implicitly labeled 'correct' only because it survived the nearest-landmark agreement criterion. This makes the statistic a self-consistency rate of the filter, not an independent measure of matching correctness. The authors should compute precision/recall against manually verified correspondences, or against 3D triangulation consistency with known bird positions, to substantiate the abstract's claim.
  2. [Abstract vs §9.1, Table 3] There is a direct numerical contradiction between the abstract's '20% elimination of outliers' and Table 3's 'Avg feature match rejection % 79.03', which is echoed in §9.1 where 'approx. 80%' of incorrect matches are expected to be removed. The paper must clarify what quantity is being reported; the current inconsistency makes the headline results ambiguous and prevents the reader from assessing the filter's effect.
  3. [§7, Eq. (1)-(2)] The load-bearing premise that the nearest landmark of a keypoint in one camera view is also its nearest landmark in the other view is asserted without sensitivity analysis. Near Voronoi cell boundaries, under occlusion, or with sparse landmark sets, perspective differences can change the nearest landmark, so true correspondences will be rejected. The authors should provide experiments varying landmark density and placement, quantify the fraction of matches near cell boundaries, and report recall of true matches in addition to the precision-like ratios in Table 3.
  4. [§9, Tables 3-5] No baseline or ablation is reported, so the claim that environmental context 'significantly improves' differentiation between visually similar birds is unsupported. The evaluation should compare the full pipeline against, at minimum, the same pipeline without the Voronoi filter, a standard ratio-test/RANSAC matching baseline, and ideally against prior work such as [23] on comparable data, using independent matching and tracking metrics.
  5. [§9.3, Table 5, Figure 6] The tracking results contradict the conclusion of 'robust and reliable tracking': Table 5 reports an average of 23.4 ID switches per minute and only 26.7% of birds tracked over 60 seconds, while Figure 6 shows 'sharp jumps' attributed to identity switching. The authors should either temper the robustness claim or provide analysis showing that the retention drop is due to unavoidable occlusion in the test intervals, with quantitative evidence.
minor comments (8)
  1. [§1.1, §7.1] The claim that 'this is the first time that Voronoi diagram is used for image tessellation' is very strong and likely incorrect; Voronoi tessellations have been used for image partitioning in prior work. Please soften the claim and cite representative earlier uses.
  2. [Eq. (2)] The typesetting of the Euclidean distance formula is corrupted (the square root symbol is missing); please fix the formula for readability.
  3. [Throughout] The repeated spacing in 'V oronoi' should be corrected to 'Voronoi' throughout the manuscript, including in section headers and Figure 5.
  4. [§3 vs Table 1] Section 3 states that each enclosure contains 20 house sparrows, but Table 1's ground-truth estimate is based on '4-5 Birds' in the test video. Please reconcile these numbers or clarify that the test clip contains only a subset of birds visible in the sampled frames.
  5. [References] Reference [8] is a duplicate of reference [7]; please remove the duplicate and renumber.
  6. [§8.3, §9.1] The Kalman filter description mentions a state of position, velocity, and acceleration, but the preceding text only discusses position and velocity; please provide the full state definition. Also, the text in §9.1 says 'ratio of correct final matches against all of the initial feature matches,' while Table 3 lists two distinct ratios; please clarify which ratio is being discussed in the text.
  7. [Table 2, Table 4] Table 2 reports keypoint statistics only for 'GoPro3 & GoPro5' even though the system uses five cameras; please explain why only two cameras are used for this table and how they are representative. Also, the evaluation intervals are described inconsistently (frames 2200-2550 in Table 2, 'three intervals of 30 frames each' in Table 4, and '3 different intervals' in §9.3); please define the experimental intervals uniformly.
  8. [Abstract] The GitHub link in the abstract is misformatted ('3D Multi Bird Tracking' lacks underscores); also, the phrase 'A viary' in the title and abstract should be 'Aviary'.

Circularity Check

1 steps flagged · score 7.0 of 10

The headline 97% matching accuracy is the Voronoi filter's own survival rate, not an independently validated accuracy, so the central claim reduces to the filter's decision rule.

  1. self definitional [Section 7 'Landmark based Outlier rejection'; Section 9.1 'Keypoints Statistics and Outlier Rejection'; Table 3]
    "Section 7: 'For a match to be valid, the nearest landmarks for the same keypoint, as observed by both paired cameras, must agree. If the nearest landmarks differ, we reject the match.' Section 9.1: 'we ... observe that the 0.97 ratio of correct final matches against all of the initial feature matches over the frame interval reflects positively on our outlier rejection methodology.' Table 3 lists 'Ratio correct final matches / all final matches 0.97'."

    The paper never defines an external ground-truth correspondence for bird keypoints. The only validity predicate in the matching pipeline is the Voronoi nearest-landmark agreement of Section 7, and 'final matches' are precisely the matches that survive that predicate. Therefore labeling a final match 'correct' and reporting 0.97 is a self-consistency rate: it restates that nearly all accepted matches satisfy the acceptance rule, not that the matches are true correspondences. The companion row 'Ratio correct final matches / all initial matches 0.20' is algebraically forced by 'Avg feature match rejection % 79.03' (1 - 0.7903 ≈ 0.21), so both headline numbers in Table 3 are functions of the filter's own accept/reject decision.

full rationale

The central empirical claim of the paper, the '97% accuracy in matching' in the abstract and conclusion, rests on Table 3's 'Ratio correct final matches / all final matches 0.97'. No ground-truth correspondences, manual verification protocol, or external benchmark is described anywhere in the paper; the only 'validity' rule in the matching stage is the Voronoi nearest-landmark agreement of Section 7, which is itself the acceptance criterion that defines which matches become final. Hence the 0.97 figure is a self-consistency rate of the filter, not a measured matching accuracy, and the 0.20 row is essentially the complement of the reported 79.03% rejection rate. This makes the abstract's headline numbers tautological with respect to the algorithm's own decision rule. The paper does contain independent components that are not circular: the YOLOv5 detection fine-tuning is compared against Mask R-CNN, the 3D reconstruction is evaluated with reprojection error in Table 4, and the tracking metrics in Table 5 are externally meaningful. Those parts could support a tracking-system contribution, but they do not validate the 'context-aware outlier rejection improves differentiation' claim because no comparison is made between the Voronoi-filtered and unfiltered matching, and no independent identity labels are used to score the matches.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The method is supported mainly by user-selected landmarks, virtual landmarks, and several untested assumptions about masks, calibration, and viewpoint invariance. No new physical entities are introduced; the virtual landmarks are algorithmic artifacts.

free parameters (5)
  • Landmark coordinates per camera view = user-defined; not reported
    The Voronoi cells are built from manually selected landmarks (Section 7), and every kept or rejected match depends on these positions. No automatic selection or sensitivity analysis is given.
  • Virtual landmarks for bounded Voronoi cells = not reported
    Section 7.1 adds virtual landmarks in a padded region to bound cells; their number and placement change cell boundaries near the image edges and thus affect matches.
  • YOLOv5 detection confidence threshold and fine-tuning hyperparameters = not reported
    Detection quality drives all downstream steps; Section 5 mentions adjusting learning rate and batch size but does not report values or the detection threshold used.
  • Kalman filter noise parameters = not reported
    Section 8.3 uses a Kalman filter with position, velocity, and acceleration states but does not specify process or measurement noise covariances, which control smoothing and association behavior.
  • Reprojection error threshold of 25 px = 25 px
    Table 4 uses 25 px to define reconstruction quality; the threshold is arbitrary and not justified, and it is applied post hoc to the collected keypoints.
assumptions (5)
  • standard math Euclidean Voronoi diagrams partition the image plane into nearest-landmark regions.
    Used in Section 7 as the basis for the outlier filter; the property is standard and cited to [19].
  • domain assumption The nearest landmark of a matched keypoint is invariant across camera views.
    Section 7 says a match is valid only if the nearest landmarks agree in both cameras. Perspective and occlusion can alter nearest-landmark identity near cell boundaries, and the paper provides no sensitivity analysis.
  • domain assumption The Canny edge and lateral-fill mask contains all relevant bird keypoints and excludes background.
    Section 6 relies on the mask to restrict SIFT keypoints to birds; if the mask leaks background or splits the bird, feature extraction and matching are degraded.
  • domain assumption The manually estimated camera calibration is accurate enough for triangulation.
    Section 8.1 describes manual intrinsic and extrinsic estimation refined by reprojection error; no independent calibration error is reported beyond the 19.8 px average reprojection error in Table 4.
  • ad hoc to paper A fixed user-defined landmark set remains valid across frames and lighting conditions.
    Section 10 states landmarks can be defined by the user; the method silently assumes these points stay fixed and identifiable throughout the recording, which is not guaranteed under occlusions or illumination changes.
invented entities (1)
  • Virtual landmarks added to the padded image region
    purpose: To bound Voronoi cells so that every pixel belongs to a finite cell, preventing infinite edges that cannot be used for segmentation.
    Section 7.1 introduces these artificial landmark points; their placement is not specified and they are not observable, but they are algorithmic constructs rather than physical entities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Context-Aware Outlier Rejection for Robust Multi-View 3D Tracking of Similar Small Birds in An Outdoor Aviary." pith.science (2026). https://pith.science/paper/EHGPRCSI

@misc{pith2026241216511,
  author       = {Pith},
  title        = {Pith review of: Context-Aware Outlier Rejection for Robust Multi-View 3D Tracking of Similar Small Birds in An Outdoor Aviary},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EHGPRCSI}},
  note         = {Machine review of arXiv:2412.16511}
}
abstract

This paper presents a novel approach for robust 3D tracking of multiple birds in an outdoor aviary using a multi-camera system. Our method addresses the challenges of visually similar birds and their rapid movements by leveraging environmental landmarks for enhanced feature matching and 3D reconstruction. In our approach, outliers are rejected based on their nearest landmark. This enables precise 3D-modeling and simultaneous tracking of multiple birds. By utilizing environmental context, our approach significantly improves the differentiation between visually similar birds, a key obstacle in existing tracking systems. Experimental results demonstrate the effectiveness of our method, showing a $20\%$ elimination of outliers in the 3D reconstruction process, with a $97\%$ accuracy in matching. This remarkable accuracy in 3D modeling translates to robust and reliable tracking of multiple birds, even in challenging outdoor conditions. Our work not only advances the field of computer vision but also provides a valuable tool for studying bird behavior and movement patterns in natural settings. We also provide a large annotated dataset of 80 birds residing in four enclosures for 20 hours of footage which provides a rich testbed for researchers in computer vision, ornithologists, and ecologists. Code and the link to the dataset is available at https://github.com/airou-lab/3D_Multi_Bird_Tracking

Figures

Figures reproduced from arXiv: 2412.16511 by the authors.

Figure 1
Figure 1. Five camera views in one of the enclosures in the aviary, which is studied in this project. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Proposed Workflow diagram: our multi-view 3D multi-bird tracking [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Masking Step. We consider masking algorithms, with the result shown in the bottom row. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 6
Figure 6. Figure 6: Graphed Trajectories of birds in Multi-view setting. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 5
Figure 5. Figure 5: Locally constructed Voronoi diagram. the plots of the tracked bird, in which in the figure there were three birds that were tracked consistently over a time interval of over a minute. Note the sharp jumps indicate the identity switching, which [PITH_FULL_IMAGE:figures…
Figure 7
Figure 7. Figure 7: Outlier rejection statistics over frame interval [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 22 canonical work pages

  1. [23]

    Multi- view tracking, re-id, and social network analysis of a flock of visually similar birds in an outdoor aviary

    Shiting Xiao, Yufu Wang, Ammon Perkes, Bernd Pfrommer, Marc Schmidt, Kostas Daniilidis, and Marc Badger. Multi- view tracking, re-id, and social network analysis of a flock of visually similar birds in an outdoor aviary. International Journal of Computer Vision, 131(6):1532–1549, 2023. 2, 3, 6 9

  2. [1]

    3D bird reconstruction: a dataset, model, and shape recovery from a single view

    Marc Badger, Yufu Wang, Adarsh Modh, Ammon Perkes, Nikos Kolotouros, Bernd Pfrommer, Marc Schmidt, and Kostas Daniilidis. 3D bird reconstruction: a dataset, model, and shape recovery from a single view. In ECCV, 2020. 2

  3. [2]

    Automated markerless pose estimation in freely moving macaques with openmonkeystudio

    Praneet C Bala, Benjamin R Eisenreich, Seng Bum Michael Yoo, Benjamin Y Hayden, Hyun Soo Park, and Jan Zim- mermann. Automated markerless pose estimation in freely moving macaques with openmonkeystudio. Nature commu- nications, 11(1):4560, 2020. 1

  4. [3]

    Measuring behaviour: an introductory guide

    Melissa Bateson and Paul Martin. Measuring behaviour: an introductory guide. Cambridge university press, 2021. 1

  5. [4]

    Learning OpenCV: Com- puter Vision with the OpenCV Library

    Gary Bradski and Adrian Kaehler. Learning OpenCV: Com- puter Vision with the OpenCV Library. O’Reilly Media, Inc.,

  6. [5]

    A computational approach to edge detection

    John Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, PAMI-8(6):91–110, 1986. 4

  7. [6]

    Pose estimation and behavior classifica- tion of broiler chickens based on deep neural networks.Com- puters and Electronics in Agriculture, 180:105863, 2021

    Cheng Fang, Tiemin Zhang, Haikun Zheng, Junduan Huang, and Kaixuan Cuan. Pose estimation and behavior classifica- tion of broiler chickens based on deep neural networks.Com- puters and Electronics in Agriculture, 180:105863, 2021. 1

  8. [7]

    Liftpose3d, a deep learning- based approach for transforming two-dimensional to three- dimensional poses in laboratory animals

    Adam Gosztolai, Semih G ¨unel, Victor Lobato-R ´ıos, Marco Pietro Abrate, Daniel Morales, Helge Rhodin, Pas- cal Fua, and Pavan Ramdya. Liftpose3d, a deep learning- based approach for transforming two-dimensional to three- dimensional poses in laboratory animals. Nature methods, 18(8):975–981, 2021. 2

Show all 23 references
  1. [8]

    Liftpose3d, a deep learning- based approach for transforming two-dimensional to three- dimensional poses in laboratory animals

    Adam Gosztolai, Semih G ¨unel, Victor Lobato-R´ıos, Marco Pietro Abrate, Daniel Morales, Helge Rhodin, Pascal Fua, and Pavan Ramdya. Liftpose3d, a deep learning- based approach for transforming two-dimensional to three- dimensional poses in laboratory animals. Nature methods, ...

  2. [9]

    Hartley and A

    R. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge books online. Cambridge Uni- versity Press, 2003. 6

  3. [10]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Gir- shick. Mask r-cnn. Proceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017. 3

  4. [11]

    Image feature match- ing and object detection using brute-force matchers

    Amila Jakubovi ´c and Jasmin Velagi´c. Image feature match- ing and object detection using brute-force matchers. In 2018 International Symposium ELMAR, pages 83–86, 2018. 5

  5. [12]

    Anipose: A toolkit for robust markerless 3d pose estimation

    Pierre Karashchuk, Katie L Rupp, Evyn S Dickinson, Sarah Walling-Bell, Elischa Sanders, Eiman Azim, Bingni W Brunton, and John C Tuthill. Anipose: A toolkit for robust markerless 3d pose estimation. Cell reports, 36(13), 2021. 1

  6. [13]

    Contributions of bird studies to biology

    Masakazu Konishi, Stephen T Emlen, Robert E Ricklefs, and John C Wingfield. Contributions of bird studies to biology. Science, 246(4929):465–472, 1989. 1

  7. [14]

    Triangulation: Why opti- mize? CoRR, abs/1907.11917, 2019

    Seong Hun Lee and Javier Civera. Triangulation: Why opti- mize? CoRR, abs/1907.11917, 2019. 6

  8. [15]

    David G. Lowe. Distinctive image features from scale- invariant keypoints. International Journal of Computer Vi- sion, 60(6):91–110, 2004. 3, 5

  9. [16]

    Markus Marks, Qiuhan Jin, Oliver Sturman, Lukas von Ziegler, Sepp Kollmorgen, Wolfger von der Behrens, Va- lerio Mante, Johannes Bohacek, and Mehmet Fatih Yanik. Deep-learning-based identification, tracking, pose estima- tion and behaviour classification of interacting primate...

  10. [17]

    Couzin, Fumihiro Kano, and Nagy M ´at´e

    Hemal Naik, Alex Hoi Hang Chan, Junran Yang, Mathilde Delacoux, Iain D. Couzin, Fumihiro Kano, and Nagy M ´at´e. 3d-pop - an automated annotation approach to facilitate markerless 2d-3d tracking of freely moving birds with marker-based motion capture. 2023 IEEE/CVF Conference ...

  11. [18]

    You only look once: Unified, real-time object de- tection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object de- tection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779–788, 2016. 3

  12. [19]

    TAMER ¨OZSU

    Cyrus Shahabi, Mehdi Sharifzadeh, LING LIU, and M. TAMER ¨OZSU. Voronoi Diagrams, pages 3438–3440. Springer US, Boston, MA, 2009. 5

  13. [20]

    Lifting multi-view detection and tracking to the bird’s eye view, 2024

    Torben Teepe, Philipp Wolters, Johannes Gilg, Fabian Her- zog, and Gerhard Rigoll. Lifting multi-view detection and tracking to the bird’s eye view, 2024. 2

  14. [21]

    3d-muppet: 3d multi-pigeon pose es- timation and tracking

    Urs Waldmann, Alex Hoi Hang Chan, Hemal Naik, M ´at´e Nagy, Iain D Couzin, Oliver Deussen, Bastian Goldluecke, and Fumihiro Kano. 3d-muppet: 3d multi-pigeon pose es- timation and tracking. International Journal of Computer Vision, pages 1–18, 2024. 2

  15. [22]

    Tracking-reconstruction or reconstruction-tracking? Technical Report BUCS-TR-2010- 030, Boston University, Department of Computer Science,

    Zheng Wu, Nickolay Hristov, Sharon Swartz, Thomas Kunz, and Margrit Betke. Tracking-reconstruction or reconstruction-tracking? Technical Report BUCS-TR-2010- 030, Boston University, Department of Computer Science,

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.