A model-agnostic active learning approach for animal detection from camera traps

Duc Thanh Nguyen; Thi Thu Thuy Nguyen

REVIEW 3 major objections 4 minor 24 references

A model-agnostic active learning approach for animal detection from camera traps

T0 review · 3 major / 4 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 A model-agnostic active learning method selects 30% of camera-trap data that trains an animal detector to equal or beat full training.

desk verdict The same-background insight and the black-box AL setup are worth attention, but the 30%-beats-full-training result is a single run with no error bars and reverses at adjacent budgets. read the letter →

arxiv 2507.06537 v1 pith:USXNLLH2 submitted 2025-07-09 cs.CV

Thi Thu Thuy Nguyen , Duc Thanh Nguyen This is my paper · ORCID

classification cs.CV

keywords activelearningmodel-agnosticanimaldetectioncameratrapsuncertaintysamplingdiversityYOLOv8SAWIT

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a model-agnostic active learning strategy can match or beat full training for animal detection from camera traps using only 30% of the labelled data. The key insight is that images from the same camera share the same background, so selecting diverse backgrounds from clusters, then picking the most uncertain image within each cluster, yields a compact but informative training set. On the SAWIT benchmark with YOLOv8, Method 1 reaches mAP@50 of 0.630 and mAP@50-95 of 0.386 at 30% data, exceeding the full-training scores of 0.611 and 0.371. If true, this would substantially cut labelling effort in wildlife monitoring.

What carries the argument

The machinery is a composite sampling score built from two quantities. Uncertainty $u(x)$ is the mean of $(1-s_i)$ over all detections in an image, $s_i$ being the detection confidence; diversity $v(x,X')$ is the mean Euclidean distance between the image's VGG-16 embedding and embeddings of a reference set. Method 1 uses K-means on embeddings to form $B$ clusters and selects the max-uncertainty image per cluster; Method 2 greedily maximizes $z(x|\hat{X}) = (1-\alpha)u(x) + \alpha v(x,\hat{X})$ with $\alpha$ decayed by the update rule. The load-bearing observation is that camera-trap frames from the same camera share almost identical backgrounds, making image-level diversity a proxy for scene coverage.

What would settle it

Re-run Method 1 on the SAWIT training pool, recording the fraction of images with $|D_f(x)|=0$ at each AL iteration, and compare the 30%-data selection when zero-detection images are either dropped entirely or assigned the maximum uncertainty; if the resulting mAP@50 no longer exceeds 0.611, or the selected set changes materially, the central claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery is that treating the target detector as a black box and combining two cheap signals—per-detection uncertainty scores and image-level diversity from a frozen pre-trained encoder—lets active learning outperform training on the entire dataset. Concretely, Method 1 partitions the unlabelled pool into B clusters by embedding distance, then selects the single highest-uncertainty image from each cluster; at 30% of the training data this exceeds full training by 1.9 points mAP@50 and 1.5 points mAP@50-95 on SAWIT. The authors interpret this as the clustering step enforcing background diversity, while the uncertainty step focuses labels where the detector is most unsure.

Load-bearing premise

The uncertainty score is undefined for images in which the detector finds no animal, so the sampling rules have nothing to rank for zero-detection images; if such images are common in the pool, the promise that 30% of the data suffices is not supported.

Editorial extensions

If this is right

If the result holds, wildlife monitoring projects can train detectors on a fraction of their labelled sets, redirecting annotation budgets to new sites or rare species.
The black-box requirement means the approach transfers to any detector that outputs boxes with confidence scores, including commercial APIs that hide model internals.
The 30%-suffices claim, if reproducible, implies that the redundancy in camera-trap datasets is large enough to be exploited without any model-specific information.
Method 1's cluster-then-select strategy suggests a general recipe: enforce diversity by partitioning, then use uncertainty to choose a representative per partition.

Reading between the lines

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

A direct extension would be to rerun Method 1 on datasets with a lower density of animals per image; one would expect the gap over full training to shrink as the fraction of zero-detection images grows.
The adaptive $\alpha$ rule in Method 2 is ad hoc; a principled scheduling based on estimated cluster sizes or validation performance might improve the later-iteration results where Method 2 currently peaks.
The embedding encoder choice is not varied; testing with a modern contrastive encoder could reveal whether the diversity term is robust to embedding quality.
Because missed detections are excluded, an easy fix is to assign zero-detection images the maximum uncertainty, which would change the selected set; comparing that variant against the reported results would quantify the sensitivity.

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.

Desk Editor's Note

read the letter

Colleague,

You should know about this paper because it attacks a real cost bottleneck in camera-trap ecology—labeling hundreds of thousands of near-duplicate background images—and it does so with a refreshingly model-agnostic setup. The core idea is to use the detector's own confidence as uncertainty and a pre-trained VGG-16 embedding as diversity, then either cluster by background and pick the most uncertain image per cluster (Method 1) or greedily trade uncertainty against novelty (Method 2). That combination is new, clearly described, and the same-background observation is genuinely useful. The authors also honestly list two limitations in the conclusion. So there is substance.

Now the soft spots. The headline claim—30% of data beats full training (0.630 vs 0.611 mAP@50)—comes from one training run with no error bars, no standard deviation, no repeated seeds. A 0.019 point gain is exactly the kind of seed-to-seed noise you see in YOLO training. Worse, the same Method 1 at 35% and 40% gives 0.593 and 0.580, below the full-training 0.611. So the "exceeds full training" is an isolated point, not a trend. That is a load-bearing weakness for the central claim.

Second, Eq. (3) is inconsistent with Algorithm 1. Eq. (3) maximizes the sum of within-cluster pairwise distances, i.e., it seeks spread-out clusters. Algorithm 1 uses K-means, which minimizes within-cluster distances. The prose says they "minimise intra-set distances," but the equation says maximize. This is a formal bug, though the algorithm itself is fine.

Third, Eq. (1) divides by |Df(x)|, so it is undefined for images with zero detections. The conclusion admits missed detections are not considered. In camera-trap data, empty frames and missed small animals are common; if those images are numerous, the sampling score gives them no ranking. This doesn't necessarily break the method, but the paper should say how often it happens and how it is dealt with.

Missing also is any comparison with [4], a camera-trap-specific AL system. It may not be directly comparable, but the reader should know why not.

These are fixable. Adding repeated runs with error bars, correcting Eq. (3), handling zero-detection images, and adding the [4] baseline would make the empirical claim supportable. As it stands, the paper is a useful idea with an unproven headline.

Who should read it: anyone working on active learning for detection, especially in camera-trap or similar redundant-background settings. It deserves a serious referee, but that referee should insist on the reruns before publication.

Referee Report

3 major / 4 minor

Summary. The paper proposes a model-agnostic active learning approach for animal detection in camera-trap images. It defines an image-level uncertainty from detection scores (Eq. (1)) and an image-level diversity from embeddings of a frozen pre-trained image encoder (Eq. (2)). Two sampling methods are presented: Method 1 partitions the unlabelled pool by K-means and selects the highest-uncertainty image per cluster, while Method 2 greedily samples using a weighted combination of uncertainty and diversity with an adaptively updated weight alpha (Eqs. (5),(6)). Experiments on the SAWIT dataset with YOLOv8 compare against random sampling, pure uncertainty, and the baselines of Roy et al. and Brust et al. The central claim is that with only 30% of the training data, Method 1 achieves mAP@50=0.630 and mAP@50-95=0.386, exceeding full-data training (0.611/0.371).

Significance. If the 30%-beats-full-training result were robust, the paper would make a practical contribution: it uses only detector outputs and a generic image encoder, and the diversity term exploits the repeated-background structure of camera-trap data. The image-level diversity idea via a task-agnostic encoder is sensible and computationally cheap as reported. The paper evaluates on a public benchmark and compares with several model-agnostic baselines. However, the evidence is currently not strong enough: the headline result is a single run with no variance estimate, the advantage at 30% is small and disappears at neighboring budgets, and no code or reproducibility statement is provided. The zero-detection limitation acknowledged in the Conclusion directly affects the core uncertainty score. With repeated runs, a corrected formalization, and a treatment of empty-detection images, the contribution could be solid.

major comments (3)

[Section 4.4, Table 1] The claim that 30% of the data suffices to exceed full-data training rests on one run without error bars. Method 1's mAP@50 at data ratios 25%, 30%, 35%, and 40% is 0.587, 0.630, 0.593, and 0.580, respectively; only the single point at 30% exceeds the full-data value of 0.611, and the gain (0.019) is within typical seed-to-seed variation for YOLO detectors. Please report means and standard deviations over at least three runs, or otherwise demonstrate that the advantage is stable and not an isolated peak.
[Section 3.2, Eq. (3) and Algorithm 1] Eq. (3) defines the 'most distinctive partition' as the partition maximizing the sum over clusters of v(x, X_k), i.e., the sum of within-cluster pairwise embedding distances. This is a scatter-maximization objective, yet Algorithm 1 obtains clusters with K-means, which minimizes within-cluster scatter. The algorithm therefore does not optimize the stated objective, and the term 'distinctive' is misleading. Please either replace Eq. (3) with the objective that Algorithm 1 actually implements (e.g., minimizing within-cluster distances) or change the algorithm to match the stated objective.
[Section 3.1, Eq. (1), and Section 5] The uncertainty score is undefined for images with no detections: u(x) divides by |Df(x)|. The Conclusion admits that missed detections are not taken into account, but this is not a minor caveat: camera-trap pools contain empty and missed-animal images, and for those images neither Algorithm 1 (step 12) nor Algorithm 2 (steps 10 and 13) can compute a score. Please state how many images in the SAWIT unlabelled pool have |Df(x)|=0, and define a fallback score (e.g., treating empty-detection images as maximally uncertain or sampling them randomly) so that the proposed methods are well-defined on the full pool.

minor comments (4)

[Section 4.4] The text reports 'mAP@50-90' in two places; this appears to be a typo for 'mAP@50-95', which is the metric defined in the same section.
[Section 3.2 / Algorithm 2 caption] The word 'unlablled' appears after Eq. (6) and should be corrected to 'unlabelled'.
[Section 3.3, Eq. (6)] The update rule alpha^(i) <- alpha^(i-1) - B/(2|X \hat X|) is not motivated, and it is unclear whether alpha is guaranteed to remain in [0,1] across iterations; please clarify or provide a rationale.
[Table 1] The caption says top-1 and top-2 performances are highlighted in bold and underline, but the table formatting does not make these highlights visible in the text; please ensure the legend is explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported active-learning gain is an empirical benchmark result with no fitted parameter renamed as a prediction.

full rationale

The central claim (30% of training data selected by Method 1 yields mAP@50 0.630 vs 0.611 for full training, Table 1) is an empirical evaluation on the SAWIT benchmark against published baselines. The uncertainty score u(x) in Eq. (1) and diversity v(x,X') in Eq. (2) are definitions used to construct the sampling criteria, not outputs derived from the final metric. The adaptive alpha in Eq. (6) is set by a hand-chosen update rule (alpha^(0)=0.5) and is not fitted to the validation mAP, so the headline result is not forced by construction. The use of the authors' own SAWIT dataset [18] for evaluation is a self-citation, but it is not load-bearing in the derivation: the dataset is an external benchmark with a fixed train/test split, and no result from [18] is invoked to justify the sampling equations. The Conclusion's caveat that 'missed detections are not taken into account' identifies a real limitation of Eq. (1) for images with |Df(x)|=0; this affects robustness and coverage but is not a circular step. Likewise, the absence of error bars and the single-run comparison concern experimental reproducibility, not circularity. No equation in the paper reduces to an input parameter renamed as a prediction, and no uniqueness theorem or ansatz is imported from self-citations. Therefore the circularity score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The method relies on two hand-set parameters (alpha_0 and the update coefficient), plus strong domain assumptions about background redundancy and embedding semantics. No new entities are introduced.

free parameters (2)

alpha_0 = 0.5
Initial weight between uncertainty and diversity in Method 2 (Eq. 5); set to a neutral 0.5 without tuning, but it is a hand-chosen constant of the algorithm.
alpha_update_coefficient = 1/2 in Eq. (6)
The factor 1/2 in the adaptive update rule for alpha in Eq. (6) is arbitrary and not derived from data or theory.

assumptions (5)

domain assumption Images from the same camera trap share similar backgrounds, so embedding-based clustering groups images by background (Section 1, Section 3.1).
This motivates image-level diversity and the K-means step; if backgrounds vary more within a camera or less across cameras, the method's advantage shrinks.
domain assumption A pre-trained ImageNet VGG16 embedding space is a suitable measure of diversity for training an animal detector (Section 4.2).
No evidence is provided that Euclidean distances in VGG16 features align with what makes camera-trap images redundant for detector training.
domain assumption The detector returns at least one detection per image, so Eq. (1) is well-defined (Eq. 1 in Section 3.1).
If an image has no detections, u(x) divides by zero; the paper lists this as a limitation in the Conclusion (missed detections are not taken into account).
ad hoc to paper K-means with K equal to the budget gives a 'most distinctive partition' as used in Algorithm 1 (Section 3.2).
The formal objective in Eq. (3) maximizes within-cluster diversity, but Algorithm 1 minimizes intra-cluster distances via K-means; the paper calls this a relaxation, which is mathematically inconsistent.
domain assumption Detector score s_i is a reliable uncertainty signal for active learning (Eq. 1).
The paper uses (1 - score) as uncertainty without calibration or analysis of the score distribution.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A model-agnostic active learning approach for animal detection from camera traps." pith.science (2026). https://pith.science/paper/USXNLLH2

@misc{pith2026250706537,
  author       = {Pith},
  title        = {Pith review of: A model-agnostic active learning approach for animal detection from camera traps},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/USXNLLH2}},
  note         = {Machine review of arXiv:2507.06537}
}

read the original abstract

Smart data selection is becoming increasingly important in data-driven machine learning. Active learning offers a promising solution by allowing machine learning models to be effectively trained with optimal data including the most informative samples from large datasets. Wildlife data captured by camera traps are excessive in volume, requiring tremendous effort in data labelling and animal detection models training. Therefore, applying active learning to optimise the amount of labelled data would be a great aid in enabling automated wildlife monitoring and conservation. However, existing active learning techniques require that a machine learning model (i.e., an object detector) be fully accessible, limiting the applicability of the techniques. In this paper, we propose a model-agnostic active learning approach for detection of animals captured by camera traps. Our approach integrates uncertainty and diversity quantities of samples at both the object-based and image-based levels into the active learning sample selection process. We validate our approach in a benchmark animal dataset. Experimental results demonstrate that, using only 30% of the training data selected by our approach, a state-of-the-art animal detector can achieve a performance of equal or greater than that with the use of the complete training dataset.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 22 canonical work pages

[4]

Dataset We validated our method on SAWIT [18], a benchmark dataset of small-sized animals captured from camera traps in the wild

EXPERIMENTS 4.1. Dataset We validated our method on SAWIT [18], a benchmark dataset of small-sized animals captured from camera traps in the wild. This dataset contains 34,434 images across seven distinct categories, including frog, lizard, bird, small mam- mal, big mammal (medium-sized mammal), spider, and scor- pion. The dataset also includes various ch...

work page
[1]

A model-agnostic active learning approach for animal detection from camera traps

INTRODUCTION Recent advances in computer vision and machine learning have created new capacities in ecological research. Auto- mated wildlife detection, empowered by deep learning, has become crucial for conservation and biodiversity assess- ments [1]. However, the performance of machine learn- ing models heavily depends on the availability of labelled da...

work page Pith review arXiv 2025
[2]

The uncertainty of a sample is often derived from the confidence score of the target model with re- spect to that sample

RELA TED WORK AL methods select samples based on two common criteria: uncertainty and diversity. The uncertainty of a sample is often derived from the confidence score of the target model with re- spect to that sample. For instance, in [6], the uncertainty of a detection was calculated from the objectness score of that de- tection and the likelihood of th...

work page
[3]

Overview Our problem of interest can be stated as follows

METHODS 3.1. Overview Our problem of interest can be stated as follows. Given a target animal detectorf(e.g., a neural network) that aims to detectCanimal classes from an input image, and a set of unlabelled animal imagesX, our objective is to select a sub- set ˆXofBimages fromXto train the detectorf. The above data sampling process can be repeated until ...

work page
[5]

This is enabled by incorporating both un- certainty and diversity quantities in the sampling process

CONCLUSION The key achievement of our AL approach is its ability to max- imise the performance of an animal detector with a signifi- cantly reduced amount of labelled data through a simple and black-box setting. This is enabled by incorporating both un- certainty and diversity quantities in the sampling process. As shown in experimental results, a target ...

work page
[6]

Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning,

Mohammad Sadegh Norouzzadeh, Anh Nguyen, Mar- garet Kosmala, Alexandra Swanson, Meredith S. Palmer, Craig Packer, and Jeff Clune, “Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning,”Proceedings of the National Academy of Sciences, vol. 115, no. 25, pp. 5716–5725, 2018

work page 2018
[7]

The iWildCam 2021 competition dataset,

Sara Beery, Arushi Agarwal, Elijah Cole, and Vighnesh Birodkar, “The iWildCam 2021 competition dataset,” CoRR, vol. abs/2105.03494, 2021

arXiv 2021
[8]

Smart camera traps and computer vision improve detections of small fauna,

Angela J. L Pestell, Anthony R Rendall, Robin D. Sin- clair, Euan G. Rictchie, Duc Thanh Nguyen, Dean M. Corva, Anne C. Eichholtzer, Abbas Z. Kouzani, and Don A. Driscoll, “Smart camera traps and computer vision improve detections of small fauna,”Ecosphere, 2025

work page 2025

Show all 24 references

[9]

A deep active learning system for species identification and counting in camera trap images,

Mohammad Sadegh Norouzzadeh, Dan Morris, Sara Beery, Neel Joshi, Nebojsa Jojic, and Jeff Clune, “A deep active learning system for species identification and counting in camera trap images,”Methods in Ecol- ogy and Evolution, vol. 12, pp. 150–161, 2019

2019
[10]

A survey on deep active learning: Recent advances and new fron- tiers,

Dongyuan Li, Zhen Wang, Yankai Chen, Renhe Jiang, Weiping Ding, and Manabu Okumura, “A survey on deep active learning: Recent advances and new fron- tiers,”IEEE Transactions on Neural Networks and Learning Systems, pp. 1–21, 2024

2024
[11]

Plug and play active learning for object detection,

Chenhongyi Yang, Lichao Huang, and Elliot J. Crowley, “Plug and play active learning for object detection,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 17784–17793

2024
[12]

Employing feature mixture for active learn- ing of object detection,

Licheng Zhang, Siew-Kei Lam, Dingsheng Luo, and Xi- hong Wu, “Employing feature mixture for active learn- ing of object detection,”Neurocomputing, vol. 594, pp. 127883, 2024

2024
[13]

SecretGen: Privacy recovery on pre-trained models via distribution discrimination,

Zhuowen Yuan, Fan Wu, Yunhui Long, Chaowei Xiao, and Bo Li, “SecretGen: Privacy recovery on pre-trained models via distribution discrimination,” inEuropean Conference on Computer Vision, 2022, pp. 139–155

2022
[14]

Simple copy-paste is a strong data augmenta- tion method for instance segmentation,

Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D. Cubuk, Quoc V . Le, and Barret Zoph, “Simple copy-paste is a strong data augmenta- tion method for instance segmentation,” inIEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2021, pp. 2918–2928

2021
[15]

Active learning for deep object de- tection,

Clemens-Alexander Brust, Christoph K ¨ading, and Joachim Denzler, “Active learning for deep object de- tection,” inInternational Joint Conference on Com- puter Vision, Imaging and Computer Graphics Theory and Applications, 2019, pp. 181–190

2019
[16]

Deep active learning for object detection.,

Soumya Roy, Asim Unmesh, and Vinay P Namboodiri, “Deep active learning for object detection.,” inBritish Machine Vision Conference, 2018, vol. 362, p. 91

2018
[17]

Instance-aware uncertainty for active learning in object detection,

Zhipeng Zhang, Wenting Ma, Xiaohang Yuan, Yuan Hao, Meng Guo, Hongyi Tang, Zhiheng Zhou, and Zhenjie Yao, “Instance-aware uncertainty for active learning in object detection,” inIEEE International Conference on Image Processing, 2024, pp. 298–304

2024
[18]

Entropy-based active learning for object detection with progressive di- versity constraint,

Jiaxi Wu, Jiaxin Chen, and Di Huang, “Entropy-based active learning for object detection with progressive di- versity constraint,” inIEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2022, pp. 9387– 9396

2022
[19]

Agnostic ac- tive learning of single index models with linear sample complexity,

Aarshvi Gajjar, Wai Ming Tai, Xingyu Xu, Chinmay Hegde, Christopher Musco, and Yi Li, “Agnostic ac- tive learning of single index models with linear sample complexity,” inAnnual Conference on Learning The- ory, Shipra Agrawal and Aaron Roth, Eds., 2024, vol. 247 ofProceedings ...

2024
[20]

Faster R-CNN: towards real-time object detection with region proposal networks,

Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun, “Faster R-CNN: towards real-time object detection with region proposal networks,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 6, pp. 1137–1149, 2017

2017
[21]

YOLO9000: better, faster, stronger,

Joseph Redmon and Ali Farhadi, “YOLO9000: better, faster, stronger,” inIEEE/CVF Computer Vision and Pattern Recognition, 2017, pp. 6517–6525

2017
[22]

Paul R Halmos,Naive set theory, Springer-Verlag, 1974

1974
[23]

SAWIT: A small-sized animal wild image dataset with annotations,

Thi Thu Thuy Nguyen, Anne C. Eichholtzer, Don A. Driscoll, Nathan I. Semianiw, Dean M. Corva, Abbas Z. Kouzani, Thanh Thi Nguyen, and Duc Thanh Nguyen, “SAWIT: A small-sized animal wild image dataset with annotations,”Multimedia Tools and Applications, vol. 83, pp. 34083–34108, 2024

2024
[24]

Ultra- lytics yolov8,

Glenn Jocher, Ayush Chaurasia, and Jing Qiu, “Ultra- lytics yolov8,” 2023

2023

Pith tools

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

[1] [4]

Dataset We validated our method on SAWIT [18], a benchmark dataset of small-sized animals captured from camera traps in the wild

EXPERIMENTS 4.1. Dataset We validated our method on SAWIT [18], a benchmark dataset of small-sized animals captured from camera traps in the wild. This dataset contains 34,434 images across seven distinct categories, including frog, lizard, bird, small mam- mal, big mammal (medium-sized mammal), spider, and scor- pion. The dataset also includes various ch...

work page

[2] [1]

A model-agnostic active learning approach for animal detection from camera traps

INTRODUCTION Recent advances in computer vision and machine learning have created new capacities in ecological research. Auto- mated wildlife detection, empowered by deep learning, has become crucial for conservation and biodiversity assess- ments [1]. However, the performance of machine learn- ing models heavily depends on the availability of labelled da...

work page Pith review arXiv 2025

[3] [2]

The uncertainty of a sample is often derived from the confidence score of the target model with re- spect to that sample

RELA TED WORK AL methods select samples based on two common criteria: uncertainty and diversity. The uncertainty of a sample is often derived from the confidence score of the target model with re- spect to that sample. For instance, in [6], the uncertainty of a detection was calculated from the objectness score of that de- tection and the likelihood of th...

work page

[4] [3]

Overview Our problem of interest can be stated as follows

METHODS 3.1. Overview Our problem of interest can be stated as follows. Given a target animal detectorf(e.g., a neural network) that aims to detectCanimal classes from an input image, and a set of unlabelled animal imagesX, our objective is to select a sub- set ˆXofBimages fromXto train the detectorf. The above data sampling process can be repeated until ...

work page

[5] [5]

This is enabled by incorporating both un- certainty and diversity quantities in the sampling process

CONCLUSION The key achievement of our AL approach is its ability to max- imise the performance of an animal detector with a signifi- cantly reduced amount of labelled data through a simple and black-box setting. This is enabled by incorporating both un- certainty and diversity quantities in the sampling process. As shown in experimental results, a target ...

work page

[6] [6]

Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning,

Mohammad Sadegh Norouzzadeh, Anh Nguyen, Mar- garet Kosmala, Alexandra Swanson, Meredith S. Palmer, Craig Packer, and Jeff Clune, “Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning,”Proceedings of the National Academy of Sciences, vol. 115, no. 25, pp. 5716–5725, 2018

work page 2018

[7] [7]

The iWildCam 2021 competition dataset,

Sara Beery, Arushi Agarwal, Elijah Cole, and Vighnesh Birodkar, “The iWildCam 2021 competition dataset,” CoRR, vol. abs/2105.03494, 2021

arXiv 2021

[8] [8]

Smart camera traps and computer vision improve detections of small fauna,

Angela J. L Pestell, Anthony R Rendall, Robin D. Sin- clair, Euan G. Rictchie, Duc Thanh Nguyen, Dean M. Corva, Anne C. Eichholtzer, Abbas Z. Kouzani, and Don A. Driscoll, “Smart camera traps and computer vision improve detections of small fauna,”Ecosphere, 2025

work page 2025