Pith. sign in

REVIEW 3 major objections 4 minor 30 references

The paper introduces PanAf-SBR, the first wild great ape camera-trap dataset with role-labelled social behaviours, and benchmarks recognition on it.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 18:03 UTC pith:37DSIQFD

load-bearing objection A genuinely useful first dataset for wild ape social behaviour, but the unvalidated labels and a duplicated table entry need fixing before the numbers can be trusted. the 3 major comments →

arxiv 2607.17399 v1 pith:37DSIQFD submitted 2026-07-19 cs.CV

The PanAf-SBR Dataset: Social Behaviour Recognition for Wild Great Apes

Maciej Braszczok , Otto Brookes , Xiaoxuan Ma , Federico Rossano , Yixin Zhu , Mimi Arandjelovic , Hjalmar K\"uhl , Majid Mirmehdi
show 1 more author
Tilo Burghardt
This is my paper
classification cs.CV
keywords great apessocial behaviour recognitioncamera trapsPanAf-SBRbehaviour benchmarktransfer learningAlphaChimpchimpanzee ethogram
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

PanAf-SBR is presented as the first wild great ape camera-trap dataset annotated with social behaviours. It extends PanAf500 with 100 videos, 36,063 frames, and 81,096 annotations covering bounding boxes, segmentation masks, intra-video identities, and seven social behaviour classes defined under the giver/receiver convention of ChimpACT. Using the AlphaChimp architecture, the paper establishes baseline benchmarks and reports that cross-dataset pre-training helps specific classes rather than all classes equally: captive ChimpACT pre-training mostly improves grooming and being_groomed, wild-to-captive pre-training mostly improves locomotion, and masking out the background lowers overall accuracy while improving contact-based social classes. A sympathetic reader would care because automated detection of social-structure breakdown could serve as an early warning for wild great ape population decline.

Core claim

The paper claims that fine-grained, role-differentiated social behaviour recognition is now possible for wild great apes in camera-trap footage. PanAf-SBR supplies the missing annotations, and the reported experiments establish three patterns: captive ChimpACT pre-training transfers mainly to grooming and being_groomed; PanAf-SBR pre-training transfers to captive locomotion classes such as sleeping, climbing, and moving; and masking out background pixels lowers mAP by 21.61 points while raising precision, recall, and F1 at the fixed threshold, with per-class gains only in contact-based social behaviours (being_carried, touch, chimp_carrying). This is the first benchmark for the task, not a s

What carries the argument

PanAf-SBR: 600 videos, 216,020 frames, 282,525 detections; each detection has a bounding box, a segmentation mask, an intra-video identity, and at least one behaviour label. Seven social labels (grooming, being groomed, chimp carrying, being carried, touch, leading, following) follow ChimpACT's giver/receiver convention, with leading and following defined independently. AlphaChimp, an end-to-end detector-tracker-classifier with a Video Swin Transformer backbone and DINO detection head, is adapted by changing the final linear layer to 16 classes. Mask-inverted videos built from the segmentation masks isolate the role of background context.

Load-bearing premise

The load-bearing premise is that the social behaviour labels are reliable: each annotation was exhaustively reviewed by a single annotator under an ethogram imported from captive chimpanzees, with no inter-annotator agreement reported, so noisy or ambiguous role assignments would propagate into every benchmark number.

What would settle it

Independently re-annotate a stratified sample of PanAf-SBR videos with at least two trained primatologists and measure agreement, especially on giver/receiver roles for grooming, carrying, leading, and following; if inter-annotator agreement falls below a conventional threshold, the benchmark numbers do not support the fine-grained social behaviour claims.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • PanAf-SBR provides the first benchmark on which wild great ape social behaviour recognition can be trained and compared, filling the gap between captive role-labelled datasets and wild footage without role labels.
  • Practitioners should expect captive pre-training to help grooming and being_groomed in the wild, while most other social classes stay flat or decline.
  • Wild pre-training should be expected to help captive locomotion recognition, especially sleeping and climbing, but not captive social interaction as a whole.
  • Background suppression is not a general accuracy booster: it lowers overall mAP across all categories, yet at a fixed threshold it raises precision/recall/F1 and specifically helps contact-based social classes.
  • Because camera locations are disjoint across train/validation/test splits, the benchmark numbers reflect site-holdout generalisation rather than per-site memorisation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The single-annotator labelling and the transfer of ChimpACT's ethogram to wild footage are the untested load-bearing parts; an inter-annotator agreement study could confirm whether the reported per-class numbers are stable.
  • The novel leading and following classes, which have no ChimpACT counterpart and stay below 30 AP, may need motion-trajectory features or clips longer than 15 seconds before they become reliable.
  • The background-masking result suggests a testable extension: use background suppression as a training-time regulariser or augmentation, not just as an input-time preprocessing step, to see whether the contact-class gains persist.
  • If social-structure breakdown is the conservation signal of interest, a follow-up could attach temporal sequence labels—who interacts with whom over minutes, not just 15-second clips—so that the benchmark speaks directly to group-level dynamics.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces PanAf-SBR, a wild great-ape camera-trap dataset that extends PanAf500 with 100 new videos, 36,063 frames, and 81,096 annotations, including bounding boxes, segmentation masks, intra-video identities, and seven social behaviour classes defined under the ChimpACT giver/receiver convention. The dataset is used with the AlphaChimp architecture to establish the first benchmarks for fine-grained social behaviour recognition in wild great apes from camera-trap footage. The paper further conducts bidirectional transfer-learning experiments between PanAf-SBR and captive ChimpACT, finding that pre-training helps specific classes (grooming, being_groomed, locomotion) rather than uniformly, and examines background removal via segmentation masks, reporting that masking lowers mAP but improves precision/recall/F1 at the chosen threshold and benefits contact-based social classes.

Significance. If the label reliability concern is adequately addressed and the reporting errors corrected, PanAf-SBR would be a valuable community resource: it is the first wild camera-trap dataset with role-differentiated social behaviour annotations, and the benchmark baselines plus transfer and background-removal analyses provide useful evidence about what does and does not transfer across captive/wild domains. The paper is transparent about several limitations, including the absence of comparative baselines and the incomplete ethogram, and it builds entirely on publicly released PanAf footage, which strengthens reproducibility. The empirical findings — that ChimpACT pre-training helps mostly grooming and being_groomed, wild-to-captive pre-training helps locomotion, and background masking improves contact-based social classes — are concrete and falsifiable. However, the absence of any inter-annotator agreement or independent ethological validation for the new social labels is a load-bearing gap, because every benchmark number inherits the label noise. The duplicated numbers in Table 4 and the omitted class in Table 2 further undermine confidence in the auditability of the reported tables.

major comments (3)
  1. [§3.3 (Data Annotation)] The social behaviour labels are the foundation of all benchmark numbers, yet no inter-annotator agreement is reported and the localisation pipeline was 'exhaustively reviewed by a single annotator'. The seven social classes are defined by transferring ChimpACT's giver/receiver convention to 15-second wild clips with no audio and often partial group context; leading and following are 'defined independently' and are the rarest classes. If role assignments are noisy or ambiguous, every AP, transfer delta, and background-masking conclusion in Tables 3–5 and Figures 5–7 inherits that error. The paper should report inter-annotator agreement (e.g., Cohen's kappa or F1 on a held-out subset) for each social label and for giver/receiver direction, and ideally have an ethologist independently validate the leading/following labels. This is a correctness risk in the ground-truth foundation, not a sty
  2. [§4.4, Table 4] The masked PanAf-SBR row reports Precision 31.95, Recall 40.77, and F1 33.21, which are numerically identical to the ChimpACT-pre-trained row in Table 3. Since background removal is a different input-space transformation, it is implausible that precision, recall, and F1 would exactly match those of the ChimpACT pre-training experiment. Please verify the numbers and rerun the evaluation if needed; as printed, the duplicated row strongly suggests a copy/paste error and undermines confidence that all benchmark tables are audited.
  3. [§4.1, Table 2] Section 4.1 states that the classification head outputs 16 classes, but Table 2 lists only 15 entries. The PanAf500 label set includes sitting_on_back (as stated in §3.3), which is missing from the mapping table. Please add sitting_on_back to Table 2 and specify its ChimpACT mapping (presumably resting). This is a concrete omission in a table that is meant to define the label space.
minor comments (4)
  1. [Figure 8] Typo: 'backgound' should be 'background'.
  2. [Tables 3–5] The paper reports point estimates without confidence intervals or error bars. Given the long-tailed class distribution and the comparatively small numbers for rare classes (leading, following), a few replications or bootstrapped confidence intervals would help readers judge whether transfer deltas are reliable.
  3. [§4.2, Fig. 6] The caption says 'Per-Class AP' but the comparison includes deltas; consider adding the baseline AP values in the figure or a supplementary table so that the reader does not need to infer them from the text.
  4. [References] ChimpACT is cited both as a dataset (ref. [19]) and through the AlphaChimp architecture (ref. [20]); the text sometimes switches between 'ChimpACT [19]' and 'ChimpACT [20]'. Please disambiguate, e.g., 'ChimpACT dataset [19]' and 'AlphaChimp [20]'.

Circularity Check

0 steps flagged

No significant circularity: the paper's benchmarks are empirical measurements; self-cited ChimpACT/AlphaChimp resources are reused tools and data, not definers of the reported results.

full rationale

The paper makes no formal derivation: it introduces a dataset, trains a model on it, and reports measured benchmark numbers. The social behaviour labels are defined following ChimpACT's giver/receiver convention, but this is an external annotation scheme, not a fitted parameter or a prediction target that is defined in terms of model outputs. The transfer-learning conclusions are empirical: e.g., the authors explicitly show that label overlap alone does not force positive transfer because touching/being_touched regress despite high pre-training frequency (Sec. 4.2), and social classes with no ChimpACT counterpart also degrade. The background-masking experiment is likewise an empirical post-hoc comparison, not a quantity that reduces to its inputs by construction. Self-citations to ChimpACT, AlphaChimp, and SA-FARI are method/tool reuse by overlapping authors, but none is a load-bearing "uniqueness theorem" invoked to forbid alternatives; the architecture and datasets are external artifacts with fixed definitions. The absence of inter-annotator agreement and reliance on a single annotator for localisation (Sec. 3.3) are data-quality limitations, not circularity. The apparent numeric duplication between Table 3's ChimpACT-pre-trained row and Table 4's masked row (31.95/40.77/33.21) is an internal reporting/auditing inconsistency to correct, not a derivation step. Under the stated rules, none of these features exhibits the required specific reduction of a claimed result to its input, so the correct finding is no significant circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

No derivation is involved. The central claim is empirical and depends on label quality, ethogram portability, and selection representativeness. These are the main uncharged assumptions; no new theoretical entities are postulated.

free parameters (2)
  • Detection and classification thresholds = IoU=0.5, classification threshold=0.5
    Evaluation working point chosen in Section 4.1; precision/recall/F1 results are threshold-dependent, and the background-removal table reports a large mAP drop together with improved P/R/F at the fixed threshold.
  • AlphaChimp hyperparameters = T=8 frames, 576x576 resolution, Q=10 queries, batch size 64, temporal stride 4
    Taken from AlphaChimp defaults in Section 4.1 and not optimized on PanAf-SBR; all benchmark numbers depend on these choices.
axioms (4)
  • domain assumption PanAf20k's video-level labels (e.g., 'social interaction', 'grooming', 'chimp carry') are sufficiently accurate to select videos containing social behaviour.
    Section 3.2 filters candidate videos based on these labels; systematic errors in upstream labels would bias the 100-video selection toward false positives or a narrow behaviour subset.
  • domain assumption ChimpACT's action-giver/receiver ethogram transfers to wild camera-trap footage without additional validation.
    Section 3.3 defines PanAf-SBR's social classes under the ChimpACT convention; wild footage has occlusion, distance, and no audio, so role assignment may be more ambiguous than in captive close-ups.
  • domain assumption SAM3 (fine-tuned on SA-FARI), plus single-annotator review and interpolation, yields correct localisation, masks, and intra-video identity tracks.
    Section 3.3 relies on this pipeline for all 600 videos; no inter-annotator agreement or error-rate measurement is reported for tracking or mask quality.
  • domain assumption The 100 added videos are representative enough of wild great ape social behavior for generalizable benchmarks.
    Section 3.2 selects for 'clearly visible, sustained social interaction', which may oversample conspicuous behaviours and exclude subtle or partially occluded interactions.

pith-pipeline@v1.3.0-alltime-deepseek · 10712 in / 14133 out tokens · 139760 ms · 2026-08-01T18:03:46.143833+00:00 · methodology

0 comments
read the original abstract

Behavioural shifts in wild great ape populations, particularly the breakdown of social structures, can serve as an early indicator of population decline. Automating the detection of behaviours indicative of these shifts is therefore a critical task for conservation. Several valuable datasets have recently been introduced for the automated recognition of great ape behaviour, yet few include fine-grained social behaviour annotations, and those that do are captured either in captive settings or via aerial platforms such as UAVs. We address this gap by introducing PanAf-SBR, the first wild great ape camera trap dataset annotated with social behaviours. PanAf-SBR extends PanAf500 with 100 additional videos covering 36,063 frames. These come with 81,096 annotations including bounding boxes, segmentation masks, intra-video identities, and seven social behaviour classes defined under the action giver and receiver convention of ChimpACT. We use this data together with the AlphaChimp architecture to establish the first benchmarks for fine-grained social behaviour recognition in wild great apes from camera trap footage. We further conduct bidirectional transfer learning experiments between PanAf-SBR and the captive ChimpACT dataset, finding that cross-dataset pre-training is highly beneficial for specific classes rather than of uniform benefit. Finally, we examine the role of background context by inverting the segmentation masks to suppress non-ape pixels.

Figures

Figures reproduced from arXiv: 2607.17399 by Federico Rossano, Hjalmar K\"uhl, Maciej Braszczok, Majid Mirmehdi, Mimi Arandjelovic, Otto Brookes, Tilo Burghardt, Xiaoxuan Ma, Yixin Zhu.

Figure 1
Figure 1. Figure 1: Overview of the PanAf-SBR Dataset. The summary illustrates the dataset and its behavioural classes including seven social behaviours defined under the action giver and receiver convention of ChimpACT. Class example visuals are shown on the left clockwise relating to the statistics on the right. All social behaviours are defined according to the ethogram used in [20], except for the leading and following, w… view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of the PanAf-SBR Behaviour Annotations. The number of samples for each behaviour category is shown. The dataset follows a typically long￾tailed distribution with many of the social behaviours falling within the middle. dataset offers the role-differentiated social annotation available in captive cor￾pora such as ChimpACT. Ape Behaviour Recognition Methods. In captive settings, pipelines in￾cre… view at source ↗
Figure 3
Figure 3. Figure 3: Data Annotations. Segmentation masks and recovered bounding boxes by the SA-FARI fine-tuned SAM3 model are shown with integers and colours encoding intra-video identity of localised animals. Frame-wise behaviour is shown as text. Bottom right image shows locomotive and social behaviour (i.e., sitting + chimp carrying). switches across frames – were corrected using purpose-built scripts for remov￾ing detect… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the Employed AlphaChimp Architecture. A Video Swin Transformer backbone, a convolutional temporal fusion layer, and a DINO detection head with a multi-label classification branch are utilised to form the AlphaChimp pipeline. (Image: courtesy of [20]). convention: every individual present in a frame receives exactly one locomotion label, with those engaged in social interaction additionally assi… view at source ↗
Figure 5
Figure 5. Figure 5: Per-category AP at IoU 0.5 for Cross-Dataset Pre-Training. Results are grouped into locomotion, object interaction, and social interaction categories. Ob￾ject interaction is the most consistent beneficiary of cross-dataset pre-training in both directions, while locomotion and social interaction transfer asymmetrically: wild-to￾captive pre-training benefits locomotion substantially but the reverse direction… view at source ↗
Figure 6
Figure 6. Figure 6: Per-Class AP on PanAf-SBR. Shown is a comparison of the baseline model against the variant pre-trained on ChimpACT. Two social behaviours (grooming and being_groomed) and one object interaction behaviour (camera_interaction) benefit from ChimpACT pre-training. The two locomotion behaviours walking and running diverge sharply: the former regresses under pre-training while the latter improves. Object Interac… view at source ↗
Figure 7
Figure 7. Figure 7: Per-class AP on ChimpACT. We compare the replicated AlphaChimp baseline against the model pre-trained on PanAf-SBR. Locomotion classes – sleeping, climbing, and moving – improve substantially, while the object interaction gain is driven almost exclusively by eating. Social interaction classes show a heterogeneous response: being_nursed regresses by −20.00%, whereas embracing improves by +36.58%. 4.4 Removi… view at source ↗
Figure 8
Figure 8. Figure 8: Input Space Background Removal. Segmentation masks are used to set backgound images to zero RGB value. Show are two ‘object only’ frames with a super￾imposition of bounding boxes and behaviours exhibited. 5 Conclusion We introduced PanAf-SBR, the first wild great ape camera trap dataset anno￾tated with social behaviours, extending PanAf500 with 100 additional videos, 36,063 frames, and 81,096 detections sp… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 2 canonical work pages

  1. [1]

    The iucn red list of threatened species.https://www.iucnredlist.org, accessed: 2026-04-30 1

    2026

  2. [2]

    de, accessed: 2026-04-30 5

    Panafricanprogramme:Theculturedchimpanzee.https://panafrican.eva.mpg. de, accessed: 2026-04-30 5

    2026

  3. [3]

    Science advances7(46), eabi4883 (2021) 4

    Bain, M., Nagrani, A., Schofield, D., Berdugo, S., Bessa, J., Owen, J., Hockings, K.J., Matsuzawa, T., Hayashi, M., Biro, D., et al.: Automated audiovisual behavior recognition in wild primates. Science advances7(46), eabi4883 (2021) 4

    2021

  4. [4]

    In: Nature Communications

    Bala, P.C., Eisenreich, B.R., Yoo, S.B.M., Hayden, B.Y., Park, H.S., Zimmermann, J.: OpenMonkeyStudio: Automated markerless pose estimation in freely moving macaques. In: Nature Communications. vol. 11, p. 4560 (2020).https://doi.org/ 10.1038/s41467-020-18441-52, 3

    work page doi:10.1038/s41467-020-18441-52 2020

  5. [5]

    arXiv preprint arXiv:2404.08471 (2024) 4

    Bardes, A., Garrido, Q., Ponce, J., Rabbat, M., LeCun, Y., Assran, M.: Revisiting feature prediction for learning visual representations from video. arXiv preprint arXiv:2404.08471 (2024) 4

    Pith/arXiv arXiv 2024

  6. [6]

    In: Proceedings- British Academy

    Boesch, C.: The emergence of cultures among wild chimpanzees. In: Proceedings- British Academy. vol. 88, pp. 251–268. OXFORD UNIVERSITY PRESS INC. (1996) 2

    1996

  7. [7]

    arXiv preprint arXiv:2301.02642 (2023) 4

    Brookes, O., Mirmehdi, M., Kühl, H., Burghardt, T.: Triple-stream deep metric learning of great ape behavioural actions. arXiv preprint arXiv:2301.02642 (2023) 4

    Pith/arXiv arXiv 2023

  8. [8]

    arXiv preprint arXiv:2404.08937 (2024) 4

    Brookes, O., Mirmehdi, M., Kuhl, H., Burghardt, T.: Chimpvlm: Ethogram- enhanced chimpanzee behaviour recognition. arXiv preprint arXiv:2404.08937 (2024) 4

    Pith/arXiv arXiv 2024

  9. [9]

    International Journal of Computer Vision132(8), 3086–3102 (2024) 2, 3, 5, 6

    Brookes, O., Mirmehdi, M., Stephens, C., Angedakin, S., Corogenes, K., Dowd, D., Dieguez, P., Hicks, T.C., Jones, S., Lee, K., et al.: Panaf20k: a large video dataset for wild ape detection and behaviour recognition. International Journal of Computer Vision132(8), 3086–3102 (2024) 2, 3, 5, 6

    2024

  10. [10]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) (2025) 2, 3

    Brookes, O., et al.: The PanAf-FGBG dataset: Understanding the impact of back- grounds in wildlife behaviour recognition. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) (2025) 2, 3

    2025

  11. [11]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3 PanAf-SBR 15

    Chen, J., et al.: MammalNet: A large-scale video benchmark for mammal recogni- tion and behaviour understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3 PanAf-SBR 15

    2023

  12. [12]

    arXiv preprint arXiv:2405.17698 (2024) 3

    Duporge, I., Kholiavchenko, M., Harel, R., Wolf, S., Rubenstein, D., et al.: Baboon- Land dataset: Tracking primates in the wild and automating behaviour recognition from drone videos. arXiv preprint arXiv:2405.17698 (2024) 3

    Pith/arXiv arXiv 2024

  13. [13]

    International Journal of Computer Vision133(10), 6668–6688 (2025) 2

    Fuchs, M., Genty, E., Bangerter, A., Zuberbühler, K., Odobez, J.M., Cotofrei, P.: From forest to zoo: great ape behavior recognition with chimpbehave. International Journal of Computer Vision133(10), 6668–6688 (2025) 2

    2025

  14. [14]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) (2023) 4

    Fuchs, M., et al.: ASBAR: An animal skeleton-based action recognition framework. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) (2023) 4

    2023

  15. [15]

    International Journal of Computer Vision (2025).https://doi.org/10

    Fuchs, M., et al.: From forest to zoo: Great ape behavior recognition with Chimp- Behave. International Journal of Computer Vision (2025).https://doi.org/10. 1007/s11263-025-02484-63

    2025

  16. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Jing, Y., Wu, J., Yang, Z., Liang, K., Zhu, X., Ma, Z.: Ethoclip: Ontology-enhanced video-language pretraining for animal behavior understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 31196–31206 (2026) 2, 3, 4

    2026

  17. [17]

    Science363(6434), 1453–1455 (2019) 1

    Kühl, H.S., Boesch, C., Kulik, L., Haas, F., Arandjelovic, M., Dieguez, P., Bocks- berger, G., McElreath, M.B., Agbor, A., Angedakin, S., et al.: Human impact erodes chimpanzee behavioral diversity. Science363(6434), 1453–1455 (2019) 1

    2019

  18. [18]

    In: Proceedings of the IEEE/CVF conference on computer vision and pat- tern recognition

    Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swin trans- former. In: Proceedings of the IEEE/CVF conference on computer vision and pat- tern recognition. pp. 3202–3211 (2022) 7

    2022

  19. [19]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 2, 3

    Ma, X., Kaufhold, S.P., Su, J., Zhu, W., Terwilliger, J., Meza, A., Zhu, Y., Rossano, F., Wang, Y.: ChimpACT: A longitudinal dataset for understanding chimpanzee behaviors. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 2, 3

    2023

  20. [20]

    arXiv preprint arXiv:2410.17136 (2024) 2, 3, 4, 6, 7, 8, 9

    Ma, X., Lin, Y., Xu, Y., Kaufhold, S.P., Terwilliger, J., Meza, A., Zhu, Y., Rossano, F., Wang, Y.: Alphachimp: Tracking and behavior recognition of chimpanzees. arXiv preprint arXiv:2410.17136 (2024) 2, 3, 4, 6, 7, 8, 9

    Pith/arXiv arXiv 2024

  21. [21]

    In: Nature Methods (2022).https://doi.org/ 10.1038/s41592-022-01396-42, 4

    Marks, M., Qureshi, T., Novak, J., Pellet, H., Issa, N., Payer, D., Wolfrum, P., Marcella, G., Contini, E., Grewe, B.F., Bohacek, J.: SIPEC: The swiss-system for primate ethology and cognition. In: Nature Methods (2022).https://doi.org/ 10.1038/s41592-022-01396-42, 4

    work page doi:10.1038/s41592-022-01396-42 2022

  22. [22]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2026), arXiv:2511.09675 2, 3, 4

    Mueller, F.B., Meier, J.F., Lueddecke, T., Vogg, R., Ecker, A.S., et al.: PriVi: To- wards a general-purpose video model for primate behavior in the wild. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2026), arXiv:2511.09675 2, 3, 4

    arXiv 2026

  23. [23]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 3

    Ng, X.L., Ong, K.E., Zheng, Q., Ni, Y., Yeo, S.Y., Liu, J.: Animal kingdom: A large and diverse dataset for animal behavior understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 3

    2022

  24. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2020) 4, 5

    Sakib, S., Burghardt, T.: Visual recognition of great ape behaviours in the wild using a dual-stream architecture. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2020) 4, 5

    2020

  25. [25]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8430–8439 (2019) 7

    2019

  26. [26]

    Nature Methods22(6), 1154–1166 (2025) 2 16 M

    Vogg, R., Lueddecke, T., Henrich, J., Dey, S., Nuske, M., Hassler, V., Murphy, D., Fischer, J., Ostner, J., Schuelke, O., et al.: Computer vision for primate behavior analysis in the wild. Nature Methods22(6), 1154–1166 (2025) 2 16 M. Braszczok et al

    2025

  27. [27]

    arXiv preprint arXiv:2511.15622 (2025) 5

    Wasmuht, D.F., Brookes, O., Schall, M., Palencia, P., Beirne, C., Burghardt, T., Mirmehdi, M., Kühl, H., Arandjelovic, M., Pottie, S., et al.: The sa-fari dataset: Segment anything in footage of animals for recognition and identification. arXiv preprint arXiv:2511.15622 (2025) 5

    arXiv 2025

  28. [28]

    Proceedings of the National Academy of Sciences114(30), 7790–7797 (2017) 2

    Whiten, A.: Culture extends the scope of evolutionary biology in the great apes. Proceedings of the National Academy of Sciences114(30), 7790–7797 (2017) 2

    2017

  29. [29]

    arXiv preprint arXiv:2203.03605 (2022) 7

    Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022) 7

    Pith/arXiv arXiv 2022

  30. [30]

    Proceedings of the IEEE109(1), 43– 76 (2020) 2

    Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., Xiong, H., He, Q.: A comprehensive survey on transfer learning. Proceedings of the IEEE109(1), 43– 76 (2020) 2

    2020