Pith. sign in

REVIEW 3 major objections 5 minor 45 references

Visual WetlandBirds Dataset: Bird Species Identification and Behavior Recognition in Videos

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The WetlandBirds dataset provides per-frame annotations of 13 bird species and 7 behaviors in 178 Spanish wetland videos, with baselines for species classification and behavior recognition.

desk verdict Useful new dataset for video bird behavior, but the missing inter-annotator agreement and an ambiguous label-priority rule keep it from being a trustworthy benchmark yet. read the letter →

arxiv 2501.08931 v1 pith:HZRW4VHL submitted 2025-01-15 cs.CV cs.AI

Javier Rodriguez-Juan , David Ortiz-Perez , Manuel Benavent-Lledo , David Mulero-Pérez , Pablo Ruiz-Ponce , Adrian Orihuela-Torres , Jose Garcia-Rodriguez , Esther Sebastián-González This is my paper · ORCID
classification cs.CVcs.AI
keywords birdbehaviorrecognitionspeciesclassificationvideodatasetwetlandbirdsspatio-temporalannotationsactioncameratrapscomputervisionbaselines
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 introduces WetlandBirds, a video dataset built from 178 recordings in Spanish wetlands, and claims it is the first fine-grained video dataset for bird behavior recognition and species classification. Existing bird datasets are mostly audio or still images, and the only prior video dataset with animal behavior annotations works at the video level for multiple animal classes. WetlandBirds instead gives per-frame bounding boxes for 13 species and 7 behaviors, organized into 2,765 behavior clips, plus fixed train/validation/test splits. The authors report baseline experiments on both tasks and interpret the modest behavior-recognition accuracy as a sign that the dataset is a real challenge for current video models. If the annotations are trustworthy, the dataset gives conservation monitoring a way to move from "which species are here" to "what are they doing."

What carries the argument

The central object is the per-frame spatio-temporal annotation format: every row in the bounding-box CSV records a global row id, video id, frame id, activity id, subject id, species, and the bounding-box coordinates, so each labeled behavior is localized in space and time. The annotation pipeline carries the argument: object-detection proposals speed up labeling, ecologist correction supplies the semantic ground truth, and Euclidean-distance matching of box centers across adjacent frames turns disconnected detections into trackable subjects. A behavior label is only applied when the behavior lasts at least 30 frames, and when a bird does several things at once the ecologically most relevant behavior is chosen, which the paper says is usually Feeding.

What would settle it

Re-annotate a random sample of the 178 videos with a second, independent team of wetland ecologists using the same seven behavior definitions and 30-frame rule, then compute Cohen's kappa per behavior. If kappa falls below 0.6 for several behaviors, the dataset's labels are not reliable enough to support the benchmark claims.

Watch

Extended reading notes

Core claim

On its own terms, the contribution is a benchmark resource rather than a new algorithm. The dataset totals 58 minutes and 53 seconds of video, with 2,765 behavior clips; each clip is a piece of video where one bird is doing one behavior for at least 30 frames. Species and behavior labels were produced by a semi-automatic workflow: ecologists assigned the species per video, an object-detection model proposed bounding boxes, ecologists corrected boxes and annotated behaviors, a Euclidean-distance rule linked boxes across frames into subject identities, and a final cross-review checked all annotations. For species classification, a YOLOv9-based baseline reaches 0.835 precision and a mean average precision (mAP50) of 0.801; for behavior recognition, a Video ResNet baseline reaches 0.56 accuracy, with MViT and SwinTransformer at 0.51 and S3D at 0.29. The authors use the gap between the two task results to argue that behavior recognition from short wetland videos remains an open problem that needs more data and better architectures.

Load-bearing premise

The dataset's ground-truth species and behavior labels are correct and consistent enough to train and evaluate models, but the paper reports no quantitative inter-annotator agreement to back this up.

Editorial extensions

If this is right

  • Automated monitoring systems could be trained to report not only which species appear in a wetland camera view, but what each bird is doing.
  • The per-frame annotations let behavior recognition be evaluated spatially and temporally, not just as a whole-video label.
  • Current video architectures reach only modest accuracy on the behavior task, so WetlandBirds is a benchmark where progress is still needed.
  • The released splits, CSV files, and code make it possible for other teams to reproduce the baselines and compare new methods.

Reading between the lines

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

  • Beyond the paper, the same annotation schema could be reused for other wetland or camera-trap sites, even though the 13 species are regional.
  • The deliberate class imbalance mirrors real recording conditions; a rebalanced version might improve accuracy metrics but would hide the difficulty of rare behaviors such as Flying and Preening.
  • An implied next step that the authors do not quantify is label reliability; adding inter-annotator agreement scores would tell users how much weight to place on each behavior class.
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 introduces WetlandBirds, a video dataset of 178 videos recorded in Spanish wetlands, annotated with 13 bird species and 7 behavior classes. The annotations are per-frame bounding boxes with subject identifiers and behavior labels, released as CSV files on Zenodo together with train/validation/test splits and code on GitHub. The authors describe a semi-automatic annotation pipeline (YOLOv7 proposals, human correction, ecologist behavior labeling, Euclidean-distance subject linking, and a final cross-review) and report baselines for species classification (YOLOv9) and behavior classification (MViT, S3D, SwinTransformer, Video ResNet). The central claim is that this is the first fine-grained video dataset designed for bird behavior detection and species classification with spatio-temporal annotations.

Significance. If the annotation quality is reliable, the dataset fills a real gap: existing video collections are either not bird-specific or do not provide per-frame, per-subject behavior labels. The authors deserve credit for releasing the data, splits, and code, and for providing baselines on two tasks. The dataset is modest in size (58 minutes, 178 videos) and strongly class-imbalanced, and the behavior baselines are near chance (accuracy 0.29--0.56), so its immediate value is as a benchmark and a starting point for method development rather than as a deployed monitoring solution. The main weakness is that the paper's central reliability claim rests on an unquantified assertion of expert annotation quality.

major comments (3)
  1. [Technical Validation] The central claim of reliable per-frame behavior and species annotations is asserted but not quantified. The Technical Validation section states that the annotators' expertise 'together with the final cross-review process, ensures the quality and cleanliness of the data,' but the paper reports no inter-annotator agreement (e.g., Cohen's kappa or per-class agreement), no count or rate of bounding-box corrections, and no measurement of how often the final cross-review changed labels. Since the dataset's value as a benchmark depends on label correctness and consistency, this is a load-bearing gap; please add a quantitative reliability study or explicitly scope the claims to avoid implying verified consistency.
  2. [Data annotation] The 'most relevant behavior' rule for simultaneous behaviors is underspecified and can create systematic label artifacts. The paragraph following Figure 1 states that when a bird performs two activities simultaneously, the most relevant behavior is annotated and Feeding always takes precedence, but no operational definition of 'most relevant' is given for other combinations and the frequency of such precedence decisions is not reported. As a result, the class Walking implicitly means 'walking without feeding,' while Feeding subsumes feeding-while-walking; models trained on these labels may learn the annotation rule rather than generalizable behavior. Please provide an explicit decision protocol, quantify how often the precedence rule was applied, or release multi-label annotations for co-occurring behaviors.
  3. [Data annotation, Subject identification] The Euclidean-distance matching between bounding-box centers of adjacent frames (Eqs. 1--2) is used to build subject tracks, but the paper does not validate this heuristic. In multi-bird videos with close or crossing individuals, nearest-center matching can swap track identities, and no identity-switch rate, manual track-correction statistics, or comparison against a reference tracking is reported. Because per-subject behavior labels are attached to these tracks, unvalidated tracking can propagate errors into the clip labels. Please quantify tracking errors or describe the manual corrections applied to tracks during data curation.
minor comments (5)
  1. [Data annotation] The text introduces 'five main steps' but then lists six items: species classification, bird localization, behavior classification, subject identification, data curation, and post-processing. Please correct the count.
  2. [Table 1] The 'Only birds' column is difficult to read and appears to contain a formatting artifact ('1pt'); also, 'BirdV ox-full-night' should be written as 'BirdVox-full-night'.
  3. [Technical Validation, Table 4] Reporting only accuracy for a highly imbalanced seven-class behavior task makes the baselines hard to interpret; please include per-class precision/recall and a confusion matrix, as is done for the species classification baseline.
  4. [Technical Validation, Behavior detection] The behavior detection baseline evaluates trimmed clips with video classification models, but the paper does not specify input resolution, number of sampled frames per clip, or how the 70-15-15 split relates to the clipped versus original videos; please add these details or point to the exact script in the released code.
  5. [Background] The sentence 'The unique dataset comprised by annotated videos with birds behavior available in the literature is the Animal Kingdom dataset' is confusing because VB100 is discussed immediately before it; please reword to make the intended contrast (bird-specific versus multi-animal, and per-frame versus video-level) explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the dataset claim, annotation process, and baseline evaluations are not derived from their own outputs.

full rationale

This is a dataset and benchmark paper, not a theoretical derivation. The central claim is the introduction of a new annotated video dataset for bird species and behavior recognition. The annotation pipeline (YOLOv7 proposals with human correction, behavior labeling by ecologists, and a final cross-review) produces ground-truth labels; it does not fit a parameter and then rename that fit as a prediction. The baseline experiments use external model architectures (YOLOv9, MViT, S3D, SwinTransformer, Video ResNet) trained on the released data, which is the standard benchmark procedure and not circular. No load-bearing self-citation chain appears: the cited prior datasets and models are external, and the paper does not invoke a uniqueness theorem or a prior same-author result to justify its design choices. The 30-frame minimum for behavior clips is a stated annotation heuristic, not a value fitted to confirm a hypothesis. The most plausible weakness is the absence of quantitative inter-annotator agreement for behavior labels and the subjective 'most relevant behavior' rule for simultaneous behaviors; however, that is a reliability/validation limitation, not a circularity in the sense of the requested analysis, because the labels are not derived from the models they are used to evaluate. Accordingly, no circular step is identified and the score is 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper relies on expert-defined behavior categories and a simple tracking heuristic, but it introduces no new physical or theoretical entities. The main load-bearing assumption is that the expert annotations are accurate and consistent, which is not quantitatively supported.

assumptions (3)
  • domain assumption The seven behavior classes (Alert, Feeding, Flying, Preening, Resting, Swimming, Walking) are mutually exclusive and exhaustively cover the relevant bird activities.
    The ecology team defined these categories based on expert knowledge (Section 'Data acquisition', Figure 1). If the taxonomy is not consistent, the annotations are hard to interpret.
  • domain assumption The 30-frame minimum duration for a behavior is a valid threshold that separates behaviors from transient sub-movements.
    This heuristic is used in the annotation process to decide when a clip represents a behavior (Section 'Background & Summary', paragraph on behavior definition).
  • ad hoc to paper The Euclidean-distance matching between bounding boxes in adjacent frames correctly tracks individual birds across time.
    This simple tracker (Equation 2) is used to build subject IDs; it can fail with overlapping or fast-moving birds, which would corrupt behavior annotations (Section 'Data annotation', step 4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Visual WetlandBirds Dataset: Bird Species Identification and Behavior Recognition in Videos." pith.science (2026). https://pith.science/paper/HZRW4VHL

@misc{pith2026250108931,
  author       = {Pith},
  title        = {Pith review of: Visual WetlandBirds Dataset: Bird Species Identification and Behavior Recognition in Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HZRW4VHL}},
  note         = {Machine review of arXiv:2501.08931}
}
read the original abstract

The current biodiversity loss crisis makes animal monitoring a relevant field of study. In light of this, data collected through monitoring can provide essential insights, and information for decision-making aimed at preserving global biodiversity. Despite the importance of such data, there is a notable scarcity of datasets featuring videos of birds, and none of the existing datasets offer detailed annotations of bird behaviors in video format. In response to this gap, our study introduces the first fine-grained video dataset specifically designed for bird behavior detection and species classification. This dataset addresses the need for comprehensive bird video datasets and provides detailed data on bird actions, facilitating the development of deep learning models to recognize these, similar to the advancements made in human action recognition. The proposed dataset comprises 178 videos recorded in Spanish wetlands, capturing 13 different bird species performing 7 distinct behavior classes. In addition, we also present baseline results using state of the art models on two tasks: bird behavior recognition and species classification.

Figures

Figures reproduced from arXiv: 2501.08931 by the authors.

Figure 1
Figure 1. Comparison of total number of clips per behavior and mean duration of behavior clips. 3/12 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Frame samples where gregarious birds appear performing different behaviors. Between the seven behaviors proposed, it should be underlined the difference between the Alert, Preening and Resting behaviors. These action distinctions were established by the ecology team. We considered that the bird was Resting when it was standing without making any movement. The bird was performing the Alert behavior when it was moving… view at source ↗
Figure 3
Figure 3. Video frame crops of bird species performing the 7 behaviors composing the dataset. 5/12 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visual representation of stages involved in the annotation process. Birds are first classified into species by annotators and localized using a YOLO model, then annotators recognize bird behaviors and subjects are identified using a Python script, and finally the data …
Figure 5
Figure 5. Figure 5: Confusion matrix of species classification pipeline. Behavior detection Secondly, the behavior detection baseline is presented. In this baseline, four different video classification models were trained end-to-end to perform the behavior classification task. The trained…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 35 canonical work pages

  1. [1]

    Environmental Science for Environmental Management(Longman, 1995)

    O’Riordan, T. Environmental Science for Environmental Management(Longman, 1995)

    work page 1995

  2. [2]

    Nichols, J. D. & Williams, B. K. Monitoring for conservation. Trends ecology & evolution 21, 668–673 (2006)

    work page 2006

  3. [3]

    Hays, G. C. et al. Translating marine animal tracking data into conservation policy and management. Trends ecology & evolution 34, 459–473 (2019)

    work page 2019

  4. [4]

    & Usher, M

    Margules, C. & Usher, M. Criteria used in assessing wildlife conservation potential: a review. Biol. conservation 21, 79–109 (1981)

    work page 1981

  5. [5]

    S., Beyea, J

    Smallwood, K. S., Beyea, J. & Morrison, M. L. Using the best scientific data for endangered species conservation. Environ. Manag. 24, 421–435 (1999)

    work page 1999

  6. [6]

    Morrison, M. L. Bird populations as indicators of environmental change. In Current Ornithology: Volume 3, 429–451 (Springer, 1986)

    work page 1986

  7. [7]

    C., Christensen, J., Boggs, C

    Bonebrake, T. C., Christensen, J., Boggs, C. L. & Ehrlich, P. R. Population decline assessment, historical baselines, and conservation. Conserv. Lett. 3, 371–378 (2010)

    work page 2010

  8. [8]

    B., Brito, J

    Carvalho, S. B., Brito, J. C., Crespo, E. J. & Possingham, H. P. From climate change predictions to actions–conserving vulnerable animal groups in hotspots at a regional scale. Glob. Chang. Biol. 16, 3257–3270 (2010)

    work page 2010

Show all 45 references
  1. [9]

    N., Maloney, R

    Joseph, L. N., Maloney, R. F. & Possingham, H. P. Optimal allocation of resources among threatened species: a project prioritization protocol. Conserv. biology 23, 328–338 (2009). 12https://pandas.pydata.org/ 13https://github.com/3dperceptionlab/Visual-WetlandBirds 10/12

    2009

  2. [10]

    Nuttall, M. N. et al. Long-term monitoring of wildlife populations for protected area management in southeast asia. Conserv. Sci. Pract. 4, e614 (2022)

    2022

  3. [11]

    Fraixedas, S. et al. A state-of-the-art review on birds as indicators of biodiversity: Advances, challenges, and future directions. Ecol. Indic. 118, 106728 (2020)

    2020

  4. [12]

    & Juranek, L

    Stastny, J., Munk, M. & Juranek, L. Automatic bird species recognition based on birds vocalization. EURASIP J. on Audio, Speech, Music. Process. 2018, 1–7 (2018)

    2018

  5. [13]

    & Planqué, R

    Vellinga, W.-P. & Planqué, R. The xeno-canto collection and its relation to sound recognition and classification. In CLEF (Working Notes) (2015)

    2015

  6. [14]

    Kahl, S. et al. Overview of birdclef 2023: Automated bird species identification in eastern africa. In CLEF (Working Notes), 1934–1942 (2023)

    2023

  7. [15]

    & Kim, A

    Miyaguchi, A., Cheung, A., Gustineli, M. & Kim, A. Transfer learning with pseudo multi-label birdcall classification for ds@gt birdclef 2024 (2024). 2407.06291

    2024 arXiv

  8. [16]

    & Stowell, D

    Morfi, V ., Bas, Y ., Pamuła, H., Glotin, H. & Stowell, D. Nips4bplus: a richly annotated birdsong audio dataset.PeerJ Comput. Sci. 5, e223 (2019)

    2019

  9. [17]

    & Bello, J

    Lostanlen, V ., Salamon, J., Farnsworth, A., Kelling, S. & Bello, J. P. Birdvox-full-night: A dataset and benchmark for avian flight call detection. In 2018 IEEE international conference on acoustics, speech and signal processing (ICASSP), 266–270 (IEEE, 2018)

    2018

  10. [18]

    & Basanta, H

    Huang, Y .-P. & Basanta, H. Bird image retrieval and recognition using a deep learning platform. IEEE Access 7, 66980–66989, 10.1109/ACCESS.2019.2918274 (2019)

    2019

  11. [19]

    & Belongie, S

    Wah, C., Branson, S., Welinder, P., Perona, P. & Belongie, S. The caltech-ucsd birds-200-2011 dataset. Tech. Rep. CNS-TR-2011-001, California Institute of Technology (2011)

    2011

  12. [20]

    Van Horn, G. et al. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 595–604, 10.1109/CVPR.2015.7298658 (2015)

    2015

  13. [21]

    L., Escos, J

    Alados, C. L., Escos, J. M. & Emlen, J. Fractal structure of sequential behaviour patterns: an indicator of stress. Animal Behav. 51, 437–443 (1996)

    1996

  14. [22]

    Lindell, C. A. The value of animal behavior in evaluations of restoration success. Restor. Ecol. 16, 197–203 (2008)

    2008

  15. [23]

    Berger-Tal, O. et al. A systematic survey of the integration of animal behavior into conservation. Conserv. Biol. 30, 744–753 (2016)

    2016

  16. [24]

    & Wittemyer, G

    Goldenberg, S., Douglas-Hamilton, I., Daballen, D. & Wittemyer, G. Challenges of using behavior to monitor anthropogenic impacts on wildlife: a case study on illegal killing of african elephants. Animal Conserv. 20, 215–224 (2017)

    2017

  17. [25]

    Deepbird: A deep learning pipeline for wildlife camera data analysis (2019)

    Harvey, S. Deepbird: A deep learning pipeline for wildlife camera data analysis (2019)

    2019

  18. [26]

    Ng, X. L. et al. Animal kingdom: A large and diverse dataset for animal behavior understanding (2022). 2204.08129

    2022 arXiv

  19. [27]

    Rose, P. et al. Evaluation of the time-activity budgets of captive ducks (anatidae) compared to wild counterparts. Appl. Animal Behav. Sci. 251, 105626 (2022)

    2022

  20. [28]

    E., Orellana, J

    Fontúrbel, F. E., Orellana, J. I., Rodríguez-Gómez, G. B., Tabilo, C. A. & Castaño-Villa, G. J. Habitat disturbance can alter forest understory bird activity patterns: A regional-scale assessment with camera-traps. For. Ecol. Manag. 479, 118618 (2021)

    2021

  21. [29]

    Fontúrbel, F. E. et al. Sampling understory birds in different habitat types using point counts and camera traps. Ecol. Indic. 119, 106863 (2020)

    2020

  22. [30]

    Murphy, A. J. et al. Using camera traps to examine distribution and occupancy trends of ground-dwelling rainforest birds in north-eastern madagascar. Bird Conserv. Int. 28, 567–580 (2018)

    2018

  23. [31]

    Arandjelovic, M. et al. Highly precise community science annotations of video camera-trapped fauna in challenging environments. Remote. Sens. Ecol. Conserv. (2024)

    2024

  24. [32]

    & Collado-Mesa, F

    Aljabri, M., AlAmir, M., AlGhamdi, M., Abdel-Mottaleb, M. & Collado-Mesa, F. Towards a better understanding of annotation tools for medical imaging: a survey. Multimed. tools applications 81, 25877–25911 (2022)

    2022

  25. [33]

    Guillermo, M. et al. Implementation of automated annotation through mask rcnn object detection model in cvat using aws ec2 instance. In 2020 IEEE REGION 10 CONFERENCE (TENCON), 708–713, 10.1109/TENCON50793.2020.9293906 (2020). 11/12

    2020

  26. [34]

    & Ogasawara, T

    Kiyokawa, T., Tomochika, K., Takamatsu, J. & Ogasawara, T. Fully automated annotation with noise-masked visual markers for deep-learning-based object detection. IEEE Robotics Autom. Lett. 4, 1972–1977, 10.1109/LRA.2019.2899153 (2019)

    2019

  27. [35]

    Pavoni, G. et al. Taglab: Ai-assisted annotation for the fast and accurate semantic segmentation of coral reef orthoimages. J. field robotics 39, 246–262 (2022)

    2022

  28. [36]

    Krenzer, A. et al. Fast machine learning annotation in the medical domain: a semi-automated video annotation tool for gastroenterologists. BioMedical Eng. OnLine 21, 33 (2022)

    2022

  29. [37]

    Li, H. et al. A semi-automated annotation algorithm based on weakly supervised learning for medical images. Biocybern. Biomed. Eng. 40, 787–802 (2020)

    2020

  30. [38]

    & Liao, H.-Y

    Wang, C.-Y ., Bochkovskiy, A. & Liao, H.-Y . M. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors (2022). 2207.02696

    2022 arXiv

  31. [39]

    & Taleb-Ahmed, A

    Sadli, R., Afkir, M., Hadid, A., Rivenq, A. & Taleb-Ahmed, A. Aggregated euclidean distances for a fast and robust real-time 3d-mot. IEEE Sensors J. 21, 21872–21884, 10.1109/JSEN.2021.3104390 (2021)

    2021

  32. [40]

    Li, A. et al. The ava-kinetics localized human actions video dataset (2020). 2005.00214

    2020 arXiv

  33. [41]

    & Liao, H.-Y

    Wang, C.-Y . & Liao, H.-Y . M. YOLOv9: Learning what you want to learn using programmable gradient information (2024)

    2024

  34. [42]

    Mvitv2: Improved multiscale vision transformers for classification and detection (2022)

    Li, Y .et al. Mvitv2: Improved multiscale vision transformers for classification and detection (2022). 2112.01526

    2022 arXiv

  35. [43]

    & Murphy, K

    Xie, S., Sun, C., Huang, J., Tu, Z. & Murphy, K. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification (2018). 1712.04851

    2018 arXiv

  36. [44]

    Liu, Z. et al. Video swin transformer (2021). 2106.13230

    2021 arXiv

  37. [45]

    CHAN-TWIN

    Tran, D. et al. A closer look at spatiotemporal convolutions for action recognition (2018). 1711.11248. Acknowledgements We would like to thank ”A way of making Europe” European Regional Development Fund (ERDF) and MCIN/AEI/10.13039/501100011033 for supporting this work under ...

    2018 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.