Pith. sign in

REVIEW 5 major objections 6 minor 59 references

Animal re-identification can run on microcontroller-class hardware with an 84 KB model.

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

2026-08-03 17:43 UTC pith:EYXCYHGC

load-bearing objection A real TinyML animal Re-ID proof-of-concept with honest limitations, but the current tables have impossible numbers and the 'practical' claim outruns the evidence. the 5 major comments →

arxiv 2512.08198 v2 pith:EYXCYHGC submitted 2025-12-09 cs.CV

Animal Re-Identification on Microcontrollers

Yubo Chen , Di Zhao , Yun Sing Koh , Talia Xu This is my paper
classification cs.CV
keywords animal re-identificationmicrocontrollerTinyMLmodel compressionlow-resolution inputMobileNetV2post-training quantizationedge inference
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.

The paper claims that animal re-identification (recognizing which individual animal a new image shows) can be done entirely on microcontroller-class hardware, not just on servers. The authors show that a carefully trimmed MobileNetV2 network—narrowed and cut to seven of its sixteen blocks—handles 64x64 images well enough to match or beat a much larger vision-language teacher on several public benchmarks, while shrinking model size by over two orders of magnitude. On a self-collected cattle dataset, an 8-bit quantized 84 KB version of the model ran fully on-device on a low-power Arduino board with only a small drop in mean average precision and unchanged top-1 accuracy compared to the floating-point cluster version. The paper also finds that knowledge distillation from large teachers gives diminishing returns under these constraints, suggesting that structural scaling is the more promising route. A reader should care because field monitoring of wildlife and livestock often happens where connectivity is unavailable, so on-device identification would remove a major bottleneck.

Core claim

At 64x64 input resolution, a MobileNetV2 backbone with width multiplier 0.35 and only the first seven of its sixteen inverted-residual blocks reaches an accuracy plateau on six public animal Re-ID benchmarks while shrinking the model from 477 MB (CLIP-ReID teacher) to 1.97 MB (FP32) and, after post-training INT8 quantization, to about 84 KB. When deployed on a low-power microcontroller (Arduino Nano 33 BLE Sense) for a self-collected cattle dataset, the quantized model performed fully on-device inference with mAP dropping from 50.1 to 45.6 and Top-1 accuracy unchanged at 61.1. The authors further claim that feature-level knowledge distillation from a ViT-based teacher (CLIP-ReID) or a CNN-ba

What carries the argument

The load-bearing mechanism is a systematic width–depth scaling of the MobileNetV2 architecture (a compact CNN built from inverted residual blocks): first fixing the width multiplier at alpha=0.35 (the narrowest ImageNet-pretrained variant, which matches wider variants at low resolution), then truncating the backbone after the first 7 of 16 bottleneck blocks because the depth-sweep shows a flat mAP plateau beyond N≈7 at 64x64. The resulting 1.97 MB FP32 model is post-training quantized to INT8 (84 KB) so it fits the microcontroller's 120 KB tensor arena. The depth-sweep experiment (Figure 4) is the key evidence for the 'knee point' claim.

Load-bearing premise

The load-bearing premise is that downsampling clean public or smartphone images to 64x64 faithfully represents what an MCU-attached camera actually captures; the paper explicitly notes (Section 6) that it does not model real camera noise, color shifts, motion blur, or limited dynamic range, and if those differ materially, the measured on-device accuracy would not transfer to the field.

What would settle it

Run the same 84 KB INT8 model on the microcontroller with a real camera module (e.g., OV7675) capturing live images of the same cattle under field conditions, and compare retrieval accuracy against the same model on downsampled clean images of the same animals. A material drop (more than a few mAP points) would falsify the claim that the reported accuracy transfers to real MCU deployment.

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

If this is right

  • Animal re-identification becomes feasible on low-power edge devices, enabling real-time monitoring in areas without internet connectivity.
  • A 477 MB server-scale model (CLIP-ReID) can be replaced by an 84 KB model with competitive or better accuracy on several datasets; on FriesianCattle2017 the compact model raises mAP by 7.9 points.
  • The finding that depth beyond ~7 blocks is redundant at 64x64 suggests that low-resolution Re-ID tasks do not need very deep networks, which could generalize to other fine-grained recognition tasks at low resolution.
  • Site adaptation can be done with as few as three images per animal by freezing the backbone and fine-tuning only the embedding layer, so deployment at new farms does not require retraining from scratch.
  • Fully on-device inference on an Arduino-class MCU is demonstrated, with a small mAP drop (50.1 to 45.6) and identical Top-1 accuracy.

Where Pith is reading between the lines

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

  • Editorial extension: The paper's evaluation uses downsampled clean images as a stand-in for real MCU camera output; a direct head-to-head test with a low-cost camera module (e.g., OV7675) would show whether the reported accuracy holds under actual sensor noise and color distortion. This is a testable extension the authors leave open.
  • Editorial extension: The depth knee at N=7 may reflect a general property of low-resolution fine-grained recognition—early layers capture the discriminative texture, and deeper layers add redundant capacity. If so, similar shallow truncations of other lightweight backbones (e.g., MobileNetV3 or EfficientNet-Lite) at 64x64 could yield comparable accuracy–size trade-offs, which is a straightforward
  • Editorial extension: The few-shot fine-tuning result (three images per identity) suggests the possibility of on-device or federated adaptation: each deployment site could fine-tune only the embedding layer with a handful of images, steadily improving recognition without any cloud round-trip. The paper does not implement on-device training, but the result makes the feasibility plausible.
  • Editorial extension: The negative distillation result could be reversed by using a stronger CNN teacher (e.g., a large ResNet or EfficientNet) instead of the few-shot teacher, which the paper did not test. This would clarify whether the architectural mismatch or the teacher's weakness is the dominant cause.

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

5 major / 6 minor

Summary. The paper proposes a TinyML pipeline for animal re-identification (Re-ID) on MCU-class devices. It first studies knowledge distillation from large ViT/CNN teachers to compact students, finding that distillation gives limited gains under MCU memory and 64x64 input constraints. It then scales a MobileNetV2 backbone by width multiplier alpha and depth N, selecting alpha=0.35 and N=7 on six public datasets, yielding a 1.97 MB FP32 model (about 84 KB INT8 after post-training quantization). The model is deployed on an Arduino Nano 33 BLE Sense with TensorFlow Lite Micro and evaluated on a self-collected cattle dataset, with the INT8 model reportedly retaining most accuracy. A few-shot fine-tuning strategy using three images per identity is also proposed. The central claim is that practical, adaptable Animal Re-ID is achievable on MCU-class devices.

Significance. If the reported numbers are correct, this is a useful feasibility demonstration: it documents that feature-level distillation is not the main lever under MCU constraints, that ImageNet initialization matters more than width at low resolution, and that a truncated, quantized MobileNetV2 can perform retrieval on a real MCU. The on-device Arduino deployment and the three-image-per-identity adaptation are concrete contributions, and the paper includes an unusually honest limitations section. However, internal inconsistencies in the headline deployment table (Table 7) and in model-size reporting, together with in-sample model selection, currently prevent full confidence in the quantitative claims. The central feasibility claim is defensible but needs corrected evidence.

major comments (5)
  1. [Table 7, Sec. 5.3] In the Arduino row, Top-5 (96.1) exceeds Top-10 (91.7). For cumulative retrieval metrics, Top-k is non-decreasing in k; Top-10 >= Top-5 always. This is an internal inconsistency in the very table used to support the abstract's claim of 'a small accuracy drop and unchanged Top-1 accuracy'. Also, the Cluster model size is listed as 1.79 MB, whereas Section 5.1 states the FP32 model is 1.97 MB. The authors must recompute and report corrected, internally consistent metrics for both the FP32 and INT8 deployments.
  2. [Secs. 4.2 - 4.3] Model-size reporting is contradictory: Section 4.2 lists the full-depth alpha=0.35 MobileNetV2 as 6.18 MB, while Section 4.3 reports the same N=16 alpha=0.35 model as 4.87 MB. The final N=7 model is 1.97 MB in Section 5.1 but 1.79 MB in Table 7, and the claimed saving is 'nearly 60%' in one paragraph and 'almost 70%' in the Section 4.3 summary (1.97/4.87 = 0.405, about 59.5%). State unambiguously which quantity (parameters, FlatBuffer bytes, or Flash footprint) is being reported and make all entries consistent.
  3. [Secs. 2.1, 5.2, 6] The abstract's 'practical ... on MCU-class devices' claim is not directly supported by the deployment experiment. Section 5.2 states that the test image is 'compiled into the firmware as constant arrays', so the input is a downsampled clean image, not a capture from the OV7675-class camera invoked in Section 2.1. Section 6 explicitly concedes that camera distortions and noise 'are not explicitly modeled in this work'. Table 7 therefore measures quantization loss on clean low-resolution images, not robustness to the field sensor pipeline. Please either scope the claims to clean low-resolution inputs or add a sensor-noise robustness study.
  4. [Secs. 4.2-4.3, Table 6] The width multiplier alpha=0.35 and depth N=7 are selected by sweeping the same six public datasets that are later used to report 'Our Method' in Table 6. The quantitative claims in the abstract ('competitive retrieval accuracy across six public datasets') are therefore post-selection estimates, which are typically optimistic. Hold out one or more datasets for selection (or use nested evaluation) and report the selection procedure; at minimum, state this in-sample caveat where Table 6 is discussed.
  5. [Sec. 2.2, Tables 1-6] The evaluation protocol for the six public datasets is under-specified: no split sizes (train/gallery/query), number of identities and queries, training epochs, triplet sampling, or margin are given, and it is unclear whether the same protocol is used across all tables. Tables 1-6 are the basis of the main accuracy claims and are not reproducible as reported. Provide a complete protocol (and ideally code) for all six datasets.
minor comments (6)
  1. [Sec. 5.1] The text says quantization 'achieves a compression ratio of over 4x' for 1.97 MB -> 84 KB, which is actually about 23x. Either the model sizes or the compression statement are wrong; please reconcile.
  2. [Sec. 4.3] Typo: 'To evalute' (should be 'evaluate'); 'leading a reduction' should be 'leading to a reduction'.
  3. [Sec. 5.3] Typo: 'recorded in the natural grazing environment of in Auckland' (remove the duplicate 'of').
  4. [Sec. 5.2] Typo: 'quantitation rule' should be 'quantization rule' (see the 'real=(q-zero_point)x scale' formula).
  5. [Table 8, Sec. 5.4] The 'No fine-tuning' row shows Top-1 72.2 versus 61.1 for 'Training from scratch', which is counterintuitive given the domain shift; add a sentence explaining this (small evaluation set size, etc.). Also clarify whether fine-tuning uses two or three images per identity.
  6. [Table 4] CLIP-ReID low-res Top-5 on ATRW (98.8) is higher than high-res (98.6) and Top-10 is equal (99.3); the text says low-resolution 'generally' lowers accuracy, which is consistent, but a brief note on this anomaly would help.

Circularity Check

0 steps flagged

No circular derivation: the architecture scaling, distillation comparisons, and MCU quantisation results are empirical and rely on external baselines; the in-sample hyperparameter selection and unmodeled camera distortions are validity caveats, not circularity.

full rationale

The paper's chain of claims is empirical rather than definitional. The width/depth study (Section 4) sweeps α and N, then measures accuracy on the six public datasets; although α=0.35 and N=7 are chosen using those same datasets and Table 6 reports performance on them, this is an in-sample hyperparameter selection bias, not a reduction of the reported accuracy to the selection criterion by construction. The knowledge-distillation analysis compares external teachers (CLIP-ReID, Few-Shot) against MobileNetV2/ViT students, and the finding that distillation offers limited benefit is an observed result, not an assumption built into the objective. The MCU deployment (Section 5) compares the FP32 cluster model with the INT8 Arduino model on the same preprocessed inputs; the small mAP drop and unchanged Top-1 are measured outcomes, not identities. The paper contains no load-bearing self-citations, no imported uniqueness theorem, and no renamed known result. Section 6 honestly states that Arduino-class camera distortions are not explicitly modeled; that is a limitation affecting real-world transfer, not circularity. The derivation is self-contained against external datasets and baselines, so no circular step is present.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central claim rests on empirical generalization from downsampled clean images to real MCU camera conditions, a proxy assumption the authors themselves flag as idealized. Architecture hyperparameters are fitted to the same benchmarks used for final reporting, and no formal or code-level support is provided.

free parameters (4)
  • MobileNetV2 width multiplier alpha = 0.35
    Selected from a width sweep (0.25-1.0) using mAP on the six public datasets; 0.35 is the narrowest ImageNet-pretrained checkpoint available (Section 4.2).
  • backbone depth N (number of retained bottleneck blocks) = 7
    Chosen as the knee point of a depth sweep N=1..16 evaluated on the same six datasets (Section 4.3, Figure 4).
  • quantization calibration set size = 100 random training images
    Used for post-training quantization calibration; no sensitivity analysis is reported (Section 5.1).
  • few-shot fine-tuning images per identity = 3
    The rapid-deployment protocol fixes three images per animal identity; no sweep is reported (Section 5.4).
axioms (6)
  • domain assumption Downsampled public or smartphone images at 64x64 are an adequate proxy for real MCU camera inputs.
    Section 2.1 sets 64x64 as the operating resolution, and Section 6 admits camera noise, color distortion, motion blur, and compression artifacts are not modeled.
  • domain assumption ImageNet-pretrained MobileNetV2 weights remain beneficial when input resolution is 64x64.
    Section 4.2 relies on Keras pretrained checkpoints for alpha >= 0.35 and shows pretraining dominates width variation.
  • domain assumption Triplet loss with L2-normalized embeddings is a suitable objective for Animal Re-ID.
    Section 4.1 replaces the softmax classifier with a metric-learning head; no comparison to other losses is provided.
  • domain assumption Gallery-query retrieval metrics (mAP, Top-k) capture real deployment performance.
    Section 2.2 defines the evaluation protocol; no live field validation is performed.
  • domain assumption The teacher models CLIP-ReID and Few-Shot are implemented and trained faithfully per their source papers.
    Section 3.2 uses external teacher models; the paper does not audit their implementations.
  • domain assumption Truncating later MobileNetV2 bottleneck blocks and attaching the original head preserves a usable embedding space.
    Section 4.3 constructs depth variants this way; the assumption is tested only indirectly through retrieval accuracy.

pith-pipeline@v1.3.0-alltime-deepseek · 19000 in / 14717 out tokens · 133989 ms · 2026-08-03T17:43:37.891185+00:00 · methodology

0 comments
read the original abstract

Camera-based animal re-identification (Animal Re-ID) can support wildlife monitoring and precision livestock management in large outdoor environments with limited wireless connectivity. In these settings, inference must run directly on collar tags or low-power edge nodes built around microcontrollers (MCUs), yet most Animal Re-ID models are designed for workstations or servers and are too large for devices with small memory and low-resolution inputs. We propose an on-device framework. First, we characterise the gap between state-of-the-art Animal Re-ID models and MCU-class hardware, showing that straightforward knowledge distillation from large teachers offers limited benefit once memory and input resolution are constrained. Second, guided by this analysis, we design a high-accuracy Animal Re-ID architecture by systematically scaling a CNN-based MobileNetV2 backbone for low-resolution inputs. Third, we evaluate the framework with a real-world dataset and introduce a data-efficient fine-tuning strategy to enable fast adaptation with just three images per animal identity at a new site. Across six public Animal Re-ID datasets, our compact model achieves competitive retrieval accuracy while reducing model size by over two orders of magnitude. On a self-collected cattle dataset, the deployed model performs fully on-device inference with only a small accuracy drop and unchanged Top-1 accuracy relative to its cluster version. We demonstrate that practical, adaptable Animal Re-ID is achievable on MCU-class devices, paving the way for scalable deployment in real field environments.

Figures

Figures reproduced from arXiv: 2512.08198 by Di Zhao, Talia Xu, Yubo Chen, Yun Sing Koh.

Figure 1
Figure 1. Figure 1: Comparison of Animal Re-Identification workflows. (a) Traditional cloud-based pipeline, where captured [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison between high-resolution and low-resolution sample images in the ATRW dataset To characterise the input quality expected within this power budget, we use the OV7675 CMOS module as an ex￾ample. While this camera is widely used with Arduino, it in￾troduces a severe quality gap compared to standard datasets: as shown in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of the MobileNetV2-based Net [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: mAP with the number of retained bottleneck [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cattle from different regions [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] 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

59 extracted references · 1 canonical work pages

  1. [1]

    Sabbir Ahmed, Abdullah Al Arafat, Deniz Najafi, Akhlak Mahmood, Mamshad Nayeem Rizve, Mohaiminul Al Nahian, Ranyang Zhou, Shaahin Angizi, and Adnan Siraj Rakin. 2025. DeepCompress-ViT: Rethinking Model Compression to Enhance Efficiency of Vision Trans- formers at the Edge. InProceedings of the Computer Vision and Pattern Recognition Conference. 30147–30156

    2025

  2. [2]

    Sungsoo Ahn, Shell Xu Hu, Andreas Damianou, Neil D Lawrence, and Zhenwen Dai. 2019. Variational information distillation for knowledge transfer. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9163–9171

    2019

  3. [3]

    Norah N Alajlan and Dina M Ibrahim. 2022. TinyML: Enabling of inference deep learning models on ultra-low-power IoT edge devices for AI applications.Micromachines13, 6 (2022), 851

    2022

  4. [4]

    William Andrew, Colin Greatwood, and Tilo Burghardt. 2017. Visual localisation and individual identification of holstein friesian cattle via deep learning. InProceedings of the IEEE international conference on computer vision workshops. 2850–2859

    2017

  5. [5]

    Sara Beery, Grant Van Horn, and Pietro Perona. 2018. Recognition in terra incognita. InProceedings of the European conference on computer vision (ECCV). 456–473

    2018

  6. [6]

    Yannick Burkard, Emanuele Francazi, Edward Lavender, Tina Dubach, Sabrina Wehrli, Jakob Brodesen, Michele Volpi, Marco Baity Jesi, and Helen Moor. 2024. Automated single species identification in camera trap images: architecture choice, training strategies, and the interpre- tation of performance metrics. (2024)

    2024

  7. [7]

    Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han

  8. [8]

    Nkosikhona Dlamini and Terence L Van Zyl. 2020. Automated iden- tification of individuals in wildlife population using siamese neural networks. In2020 7th international conference on soft computing & machine intelligence (ISCMI). IEEE, 224–228

    2020

  9. [9]

    Alexey Dosovitskiy. 2020. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2020

  10. [10]

    Jin Gao, Shubo Lin, Shaoru Wang, Yutong Kou, Zeming Li, Liang Li, Congxuan Zhang, Xiaoqin Zhang, Yizheng Wang, and Weiming Hu

  11. [11]

    Google AI Edge Team. 2024. Post-training quantization. Google AI Edge LiteRT Documentation. https://ai.google.dev/edge/litert/models/ post_training_quantization Accessed: 2025-11-21

    2024

  12. [12]

    Xi Guo, Yufeng Chen, Yu Guan, Hongfang Wang, Tianming Wang, Jianping Ge, and Lei Bao. 2025. Individual identification of wild raptors using a deep learning approach: A case study of the white-tailed eagle. Ecological Informatics(2025), 103379

    2025

  13. [13]

    Zhimin He, Jiangbo Qian, Diqun Yan, Chong Wang, and Yu Xin. 2023. Animal re-identification algorithm for posture diversity. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1–5

    2023

  14. [14]

    Soroush Heydari and Qusay H Mahmoud. 2025. Tiny machine learning and on-device inference: A survey of applications, challenges, and future directions.Sensors25, 10 (2025), 3191

    2025

  15. [15]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531(2015)

    Pith/arXiv arXiv 2015

  16. [16]

    Md Imtiaz Hossain, Sharmen Akhter, Choong Seon Hong, and Eui-Nam Huh. 2025. Single teacher, multiple perspectives: Teacher knowledge augmentation for enhanced knowledge distillation. InThe Thirteenth International Conference on Learning Representations

    2025

  17. [17]

    Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam

  18. [18]

    Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko

  19. [19]

    Jiening Jiao, Wei-Shi Zheng, Ancong Wu, Xiatian Zhu, and Shaogang Gong. 2018. Deep low-resolution person re-identification. InProceed- ings of the AAAI conference on artificial intelligence, Vol. 32

    2018

  20. [20]

    Xin Jin, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. 2020. Uncertainty- aware multi-shot knowledge distillation for image-based object re- identification. InProceedings of the AAAI Conference on Artificial Intel- ligence, Vol. 34. 11165–11172

    2020

  21. [21]

    Xiao-Yuan Jing, Xiaoke Zhu, Fei Wu, Xinge You, Qinglong Liu, Dong Yue, Ruimin Hu, and Baowen Xu. 2015. Super-resolution person re- identification with semi-coupled low-rank discriminant dictionary learning. InProceedings of the IEEE conference on computer vision and pattern recognition. 695–704

    2015

  22. [22]

    Emmanuel Kabuga, Izzy Langley, Monica Arso Civil, John Measey, Bubacarr Bah, and Ian Durbach. 2024. Similarity learning networks uniquely identify individuals of four marine and terrestrial species. Ecosphere15, 10 (2024), e70012. doi:10.1002/ecs2.70012

    work page doi:10.1002/ecs2.70012 2024

  23. [23]

    Taylor L Kaltenbach, Jeffrey C Mosley, Lance B McNew, and Jared T Beaver. 2025. Can edge AI mitigate environmental effects on camera trap performance?Wildlife Society Bulletin(2025), e1598

    2025

  24. [24]

    Zhenglun Kong, Dongkuan Xu, Zhengang Li, Peiyan Dong, Hao Tang, Yanzhi Wang, and Subhabrata Mukherjee. 2025. AutoViT: Achieving Real-Time Vision Transformers on Mobile via Latency-aware Coarse- to-Fine Search.International Journal of Computer Vision(2025), 1–17

    2025

  25. [25]

    Shuyuan Li, Jianguo Li, Hanlin Tang, Rui Qian, and Weiyao Lin. 2019. ATRW: a benchmark for Amur tiger re-identification in the wild.arXiv preprint arXiv:1906.05586(2019)

    Pith/arXiv arXiv 2019

  26. [26]

    Siyuan Li, Li Sun, and Qingli Li. 2023. Clip-reid: exploiting vision- language model for image re-identification without concrete text labels. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 1405–1413

    2023

  27. [27]

    Ji Lin, Wei-Ming Chen, Yujun Lin, Chuang Gan, Song Han, et al. 2020. Mcunet: Tiny deep learning on iot devices.Advances in neural infor- mation processing systems33 (2020), 11711–11722

    2020

  28. [28]

    Amir Moslemi, Anna Briskina, Zubeka Dang, and Jason Li. 2024. A survey on knowledge distillation: Recent advancements.Machine Learning with Applications18 (2024), 100605

    2024

  29. [29]

    Margarita Mulero-Pázmány, Sandro Hurtado, Cristóbal Barba- González, María Luisa Antequera-Gómez, Francisco Díaz-Ruiz, Raimundo Real, Ismael Navas-Delgado, and José F Aldana-Montes

  30. [30]

    Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bon- darenko, Mart Van Baalen, and Tijmen Blankevoort. 2021. A white paper on neural network quantization.arXiv preprint arXiv:2106.08295 (2021)

    Pith/arXiv arXiv 2021

  31. [31]

    Maarten Perneel, Ines Adriaens, Jan Verwaeren, and Ben Aernouts

  32. [32]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural Conf ’25, Date, Location Yubo Chen, Di Zhao, Yun Sing Koh, and Talia Xu language supervision. InInternational conference on machine learning. PmLR,...

    2021

  33. [33]

    Mohammad Mehdi Rastikerdar, Jin Huang, Hui Guan, and Deepak Ganesan. 2024. In-Situ Fine-Tuning of Wildlife Models in IoT-Enabled Camera Traps for Efficient Adaptation.arXiv preprint arXiv:2409.07796 (2024)

    arXiv 2024

  34. [34]

    Addressing significant challenges for animal detection in camera trap images: a novel deep learning-based approach.Scientific Reports 15, 1 (2025), 16191

    2025

  35. [35]

    Shaibal Saha and Lanyu Xu. 2025. Vision transformers on the edge: A comprehensive survey of model compression and acceleration strate- gies.Neurocomputing(2025), 130417

    2025

  36. [36]

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. 2018. Mobilenetv2: Inverted residuals and linear bottlenecks. InProceedings of the IEEE conference on computer vision and pattern recognition. 4510–4520

    2018

  37. [37]

    Dynamic Multi-Behaviour, Orientation-Invariant Re- Identification of Holstein-Friesian Cattle.Sensors25, 10 (2025), 2971

    2025

  38. [38]

    Mingxing Tan and Quoc Le. 2019. Efficientnet: Rethinking model scaling for convolutional neural networks. InInternational conference on machine learning. PMLR, 6105–6114

    2019

  39. [39]

    Keras Team. [n. d.]. MobileNet, MobileNetV2, and MobileNetV3. Keras 3 API documentation. https://keras.io/api/applications/mobilenet/ Accessed: 2025-11-21

    2025

  40. [40]

    Partha Pratim Ray. 2022. A review on TinyML:: State-of-the-art and prospects. (2022)

    2022

  41. [41]

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. 2021. Training data-efficient image transformers & distillation through attention. InInternational conference on machine learning. PMLR, 10347–10357

    2021

  42. [42]

    Devis Tuia, Benjamin Kellenberger, Sara Beery, Blair R Costelloe, Silvia Zuffi, Benjamin Risse, Alexander Mathis, Mackenzie W Mathis, Frank Van Langevelde, Tilo Burghardt, et al. 2022. Perspectives in machine learning for wildlife conservation.Nature communications13, 1 (2022), 792

    2022

  43. [43]

    Moritz Scherer, Fabian Sidler, Michael Rogenmoser, Michele Magno, and Luca Benini. 2022. Widevision: A low-power, multi-protocol wireless vision platform for distributed surveillance. In2022 18th In- ternational Conference on Wireless and Mobile Computing, Networking and Communications (WiMob). IEEE, 394–399

    2022

  44. [44]

    Oscar Wahltinez and Sarah J Wahltinez. 2024. An open-source gen- eral purpose machine learning framework for individual animal re- identification using few-shot learning.Methods in Ecology and Evolu- tion15, 2 (2024), 373–387

    2024

  45. [45]

    Le Wang, Rizhi Ding, Yuanhao Zhai, Qilin Zhang, Wei Tang, Nan- ning Zheng, and Gang Hua. 2021. Giant panda identification.IEEE Transactions on Image Processing30 (2021), 2837–2849

    2021

  46. [46]

    TensorFlow Authors. 2022. Post-training quantization. TensorFlow Model Optimization Guide. https://www.tensorflow.org/model_ optimization/guide/quantization/post_training Accessed: 2025-11-21

    2022

  47. [47]

    Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Yanghan Wang, Fei Sun, Yiming Wu, Yuandong Tian, Peter Vajda, Yangqing Jia, and Kurt Keutzer. 2019. Fbnet: Hardware-aware efficient convnet design via dif- ferentiable neural architecture search. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10734–10742

    2019

  48. [48]

    Zhendong Yang, Zhe Li, Ailing Zeng, Zexian Li, Chun Yuan, and Yu Li. 2024. Vitkd: Feature-based knowledge distillation for vision trans- formers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 1379–1388

    2024

  49. [49]

    Delia Velasco-Montero, Jorge Fernández-Berni, Ricardo Carmona- Galán, Ariadna Sanglas, and Francisco Palomares. 2024. Reliable and efficient integration of AI into camera traps for smart wildlife moni- toring based on continual learning.Ecological Informatics83 (2024), 102815

    2024

  50. [50]

    Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. 2015. Scalable person re-identification: A benchmark. InProceedings of the IEEE international conference on computer vision. 1116–1124

    2015

  51. [51]

    Yujie Zhong, Xiao Li, Jiangjian Xie, and Junguo Zhang. 2023. A light- weight automatic wildlife recognition model design method mitigating shortcut learning.Animals13, 5 (2023), 838

    2023

  52. [52]

    Ancong Wu, Wei-Shi Zheng, Xiaowei Guo, and Jian-Huang Lai. 2019. Distilled person re-identification: Towards a more scalable system. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 1187–1196

    2019

  53. [55]

    Guoqing Zhang, Yuhao Chen, Weisi Lin, Arun Chandran, and Xuan Jing. 2021. Low resolution information also matters: Learning multi- resolution representations for person re-identification.arXiv preprint arXiv:2105.12684(2021)

    Pith/arXiv arXiv 2021

  54. [58]

    Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy

  55. [2017]

    Mobilenets: Efficient convolutional neural networks for mobile vision applications.arXiv preprint arXiv:1704.04861(2017)

    Pith/arXiv arXiv 2017

  56. [2018]

    InProceedings of the IEEE conference on computer vision and pattern recognition

    Quantization and training of neural networks for efficient integer- arithmetic-only inference. InProceedings of the IEEE conference on computer vision and pattern recognition. 2704–2713

  57. [2019]

    Once-for-all: Train one network and specialize it for efficient deployment.arXiv preprint arXiv:1908.09791(2019)

    Pith/arXiv arXiv 1908

  58. [2022]

    Domain generalization: A survey.IEEE transactions on pattern analysis and machine intelligence45, 4 (2022), 4396–4415

    2022

  59. [2025]

    An Experimental Study on Exploring Strong Lightweight Vision Transformers via Masked Image Modeling Pre-training.International Journal of Computer Vision133, 7 (2025), 3918–3950

    2025