Pith. sign in

REVIEW 4 major objections 8 minor 34 references

Always-on low-power sentinels plus a rarely woken Jetson can turn raw underwater detections into local taxonomic knowledge and reports without continuous GPU power or raw-data uplink.

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 · grok-4.5

2026-07-31 18:24 UTC pith:AYDL7BME

load-bearing objection Solid energy-cascade systems paper with real hardware numbers; the species-ID half of the headline claim is undercut by the authors’ own OzFish results. the 4 major comments →

arxiv 2607.24313 v1 pith:AYDL7BME submitted 2026-07-27 cs.IR cs.CVcs.MAcs.RO

Energy Constrained Hierarchical Underwater Monitoring via Local Multi-Agent RAG

Mohamed Amine Janati , Laurent Gautier , St\'ephane Barbot This is my paper
classification cs.IR cs.CVcs.MAcs.RO
keywords Ocean observationLocal AIAgentic AIRetrieval-Augmented GenerationSpecies classificationEmbedded AIHierarchical edge sensingUnderwater monitoring
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.

Underwater monitoring stations face tight batteries, weak links, and costly raw multimodal uploads, so most data sit on disk until recovery. This paper argues that a hierarchical master–satellite design solves that: ultra-low-power MAX78000/MAX78002 chips run continuous visual and acoustic event filters, while an NVIDIA Jetson Orin NX stays off until a schedule, a sentinel trigger, or a researcher needs it. Once awake, the Jetson runs a fully local pipeline—crop extraction, BioCLIP/OpenCLIP embeddings into ChromaDB, centroid and SVM species hypotheses, multi-agent retrieval-augmented reasoning, and Markdown reporting—so the station emits structured, scientist-ready summaries instead of video and audio dumps. The claim is that this cascade keeps continuous sensing cheap, reserves heavy compute for justified moments, and compresses what must be transmitted over acoustic, optical, or satellite links. A sympathetic reader cares because the same pattern could let remote marine stations stay out for months and still answer taxonomic and temporal questions at the edge.

Core claim

A hierarchical energy-aware architecture that keeps MAX78000/MAX78002 visual and acoustic sentinels always on and activates a Jetson Orin NX only for scheduled, event-driven, or interactive work can run a fully local multimodal pipeline—embedding, adaptive species identification, multi-agent RAG, and reporting—thereby producing researcher-ready knowledge while cutting energy use and communication volume relative to continuous high-performance sensing and raw-data uplink.

What carries the argument

The hierarchical master–satellite cascade: always-on microcontroller sentinels filter events; a selectively powered Jetson runs BioCLIP/OpenCLIP embeddings, a dedicated identification layer (centroids plus SVMs in embedding space), local ChromaDB collections, and a LangChain multi-agent controller (router, analyst, energy, hardware, reporting) that grounds answers and reports in indexed mission and taxonomic data.

Load-bearing premise

That sentinel accuracy and confidence thresholds will keep Jetson wake-ups rare enough, and that the measured per-block energy model with assumed daily frequencies will translate into multi-month battery life under real turbidity, fouling, and noise.

What would settle it

A closed-loop multi-month battery deployment that logs actual sentinel false-positive and true-event rates, realized Jetson daily energy against the block model, and whether the local reports remain taxonomically usable under field imaging and acoustic conditions.

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

If this is right

  • Remote stations can store and later uplink compact embeddings and Markdown summaries instead of raw video and audio.
  • Reference collections can be swapped per mission (invasive species, benthos, mammals) without retraining the backbone encoders.
  • Three modes—autonomous batch, event-driven hybrid, and researcher-interactive—can share one hardware stack across long deployments and field visits.
  • Energy and hardware agents can retune FPS, wake thresholds, and microcontroller model binaries from remaining battery and mission priorities.
  • The same cascade pattern extends to ROV/AUV payloads and solar buoys that need edge interpretation under power limits.

Where Pith is reading between the lines

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

  • If wake rates stay low in the field, the architecture becomes a template for other bandwidth-starved domains (glacier cams, forest acoustics) that need local RAG without always-on GPUs.
  • Cross-dataset species drops on real-condition sets like OzFish imply that reference curation and hierarchical fallback matter as much as the cascade itself for trustworthy edge labels.
  • A projector that feeds stored CLIP vectors straight into a small local LLM, as the authors sketch in future work, would further cut the cost of report generation once the identification layer is trusted.

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

4 major / 8 minor

Summary. The manuscript presents a hierarchical underwater monitoring architecture in which always-on MAX78000 (acoustic) and MAX78002 (visual) microcontroller "sentinels" perform first-stage detection and selectively wake a Jetson Orin NX that runs a fully local multimodal pipeline: Grounding DINO/YOLO-World localization, BioCLIP-2/OpenCLIP embedding into ChromaDB, a centroid+SVM species identification layer over five taxonomic reference collections, a LangChain multi-agent framework (Router, Analyst, Energy/Sensor, Model Deployment, Reporting), and Markdown report generation for low-bandwidth uplink. Three operating modes (batch, event-driven, interactive) are supported by a hardware wake/power-hold circuit. Evaluation covers sentinel accuracy (visual mAP@50 70.75–78.4% on filtered FishDet-M; acoustic ~99% accuracy on a 13-class Watkins/DeepShip task), cross- and intra-dataset taxonomic identification (Tables 5–6), and block-wise measured energy consumption on both tiers (Tables 7–9), with a block-additive daily energy model (Eqs. 18–20).

Significance. If the claims are scoped to what is demonstrated, this is a useful and honest systems contribution: the hierarchical MAX78000/MAX78002-plus-Jetson design with a working hardware wake/hold circuit (Fig. 4) is a credible pattern for long-duration underwater stations; the sentinel metrics on external benchmarks (Table 4), the nine-block instrumented energy characterization (Tables 7–9), and the taxonomic evaluation against a deliberately held-out real-world set (OzFish) are concrete, reproducible artifacts. Notably, the paper reports its own negative cross-domain result (Table 5) rather than hiding it, which is to the authors' credit — but they must also draw the consequences of it in the claims and discussion. The work is an architecture/feasibility study rather than a validated deployment, and publication is warranted once the presentation matches that status.

major comments (4)
  1. [§6.3, Table 5] Table 5 (OzFish column) vs. §1/§4.3.4/abstract: on the only real in-situ evaluation set, the identification layer collapses — species Top-1 is 0.7–4.2% (Top-5 ≤ 8.2%) for every reference collection, and the hierarchical fallback the paper relies on (§4.3.4: 'a broader taxonomic label is retained instead of forcing an unreliable species assignment') also fails: family Top-1 is only 13.4–16.7% and genus ≤ 13.9%. Since §4.3.3 stores these species hypotheses as metadata and §4.3.5 groups retrieval by 'reliable' BioCLIP scores, the downstream RAG layer grounds its reports on labels that are wrong at species level ~96% of the time under deployment-realistic conditions. This directly undercuts the 'researcher-ready structured knowledge' claim (abstract, §1) and the 'dedicated species identification layer' contribution. Yet §7 (Discussion) never mentions Table 5's OzFish results. The authors sho
  2. [§6.3.1, Eqs. (18)–(20)] Eqs. (18)–(20) and Table 9: the energy-saving half of the central claim is never instantiated. The paper decomposes Jetson operation into nine measured blocks but never computes E_total for any concrete scenario — the daily frequencies n_i / λ_event are left 'hypothetical', and no comparison against an always-on Jetson baseline (the quantity the architecture exists to reduce) is given anywhere. Moreover, the assumption under Eq. (20) that 'the time between operations is negligible' is questionable given the paper's own Table 9: sentinel-data transfer takes 136 s for only 1 MB at 115200 baud (0.26 Wh, the second-largest block), so for a day of accumulated detections the transfer block could dominate and the inter-block idle gaps are not obviously negligible. A worked end-to-end example — e.g., one daily wake cycle with a stated detection volume, total Wh/day for each of the three modes ve
  3. [§3.2, §6.2, Table 4] Table 4, acoustic row: the reported 98.99% accuracy / 96.4% F1 carries a correctness risk from the split protocol. §3.2 states a 70/20/10 split of the aggregated dataset, but does not say whether the split is at the recording-file/session level or at the 0.98 s window level. If windows from the same Watkins recording appear in both train and test, the figure is inflated by near-duplicate leakage. In addition, the Noise class is built from DeepShip (ship recordings) plus synthetic board noise — not ambient ocean sound — so the model's ability to reject real underwater backgrounds (the property that determines false Jetson wake-ups in Hybrid Mode, §4.4.3) is untested. Concrete fixes: state the split granularity; re-evaluate with sessions held out; and report a false-accept rate on a held-out ambient-noise corpus, since this number, not clean-test accuracy, drives the event-driven energy bu
  4. [§6.3 (agentic evaluation)] The multi-agent control framework — a title contribution with five agents (§4.4) — is evaluated only by qualitative 'functional validation' (§6.3). The only quantitative statements are anecdotal (Qwen3.5-4B 'frequently required up to eight iterations'; Ornith-9B correct 'on the first iteration in most cases'), with no task suite, sample sizes, routing-accuracy numbers, or report-quality checks. The authors are candid about this limitation, which counts in their favor, but for a cs.IR venue the retrieval/routing layer in particular needs some measurable evaluation: even a small fixed query set (e.g., 30–50 researcher queries spanning the four retrieval modalities and the analytical fast path) with routing accuracy, answer groundedness against cited sources, and Analyst execution success rate would make this contribution assessable rather than asserted.
minor comments (8)
  1. [Table 5] Table 5 lacks confidence intervals or per-cell sample counts. OzFish has only 1,903 images across 198 species; species-level Top-1 of 3.2% vs 4.2% may be within noise. Please report binomial CIs or per-class support.
  2. [Table 9] Table 9, block 2: the transfer workload (1 MB, 32 audio files) seems unrepresentative of a full day of visual sentinel detections; please state expected daily data volume per sentinel and scale the block accordingly in any scenario calculation.
  3. [§4.4.2, §6.3] The models Qwen3.5-4B and Ornith-9B are introduced without citation, version, or quantization details. 'Ornith-9B' in particular is not identifiable from the text; reproducibility requires exact model identifiers and Ollama/quantization configurations.
  4. [§4.3.5] Key configuration choices are stated but not ablated or justified: τ_g = 0.86 (Eq. 17), the 6:4 visual:textual retrieval mix, the 4000-character context cap, and the top-N crop retention. A brief sensitivity note would suffice.
  5. [Table 7] Table 7 header has a formatting error ('FPS ms' appears as merged/duplicated column labels); also clarify that FPS and latency are separate columns.
  6. [Various] Typos and wording: 'mentionned' (§4.1), 'severals months' (§1), 'Rasberry Pi' (§2.1), 'specie identification' (Fig. 5 caption and §4.2.3 — should be 'species'). Figure 5's caption says it shows five taxonomic levels while §4.3.4 says it displays class level only; please reconcile.
  7. [§6.2] The visual sentinel's mAP@50 (70.75–78.4%) is measured on FishDet-M filtered to ≤5 boxes per image; please note explicitly in §6.2 that this is an easier subset than the full benchmark (Table 1 shows the filtering removes ~60% of instances), so readers do not compare these numbers to full-FishDet-M results.
  8. [End of manuscript] No code, model, or data availability statement is given. Given the engineering nature of the contribution, releasing the block-measurement scripts, the wake-circuit schematic details, and the reference-collection index manifests would substantially increase the paper's value.

Circularity Check

0 steps flagged

No significant circularity: engineering system evaluated on external datasets and measured hardware energy blocks, not a self-referential derivation.

full rationale

This paper proposes and characterizes a hierarchical underwater monitoring architecture (MAX78000/02 sentinels + selectively woken Jetson Orin NX with local multimodal RAG). Its load-bearing claims are empirical and architectural: sentinel mAP/accuracy on FishDet-M, Watkins, and DeepShip; cross- and intra-dataset taxonomic identification on FishBase, WildFish++, Fish-Vista, FishNet, BioTrove, OzFish; and discrete Wh measurements of boot, transfer, indexing, RAG, and shutdown blocks. The energy model (Eqs. 18–20) is ordinary additive accounting—E_total = E_sentinel + Σ n_i E_b_i—with n_i left as mode-dependent or hypothetical frequencies, not fitted then re-presented as independent predictions. Configuration choices (τ_g = 0.86, 6:4 visual/text retrieval mix, confidence thresholds) are stated as deployment parameters, not derived uniqueness results. The sole self-citation to prior Lionfish/Raspberry Pi work is background related work and is not used to force the architecture or forbid alternatives. There is no self-definitional loop, no fitted-input-called-prediction, no load-bearing uniqueness import, and no renaming of a known law as a new derivation. Poor OzFish species Top-1 in Table 5 is a correctness/generalization failure, not circularity. Score 0 is appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

Load-bearing content is systems composition plus empirical thresholds and energy block measurements. Claims rest on domain hardware limits, embedding-space taxonomy assumptions, and unvalidated long-horizon event-rate/energy extrapolations rather than free-form physical entities.

free parameters (5)
  • Visual grouping similarity threshold τ_g = 0.86
    Default 0.86 used to cluster unlabeled crops within a video before RAG context packing; chosen configuration, affects report compression and evidence diversity.
  • Retrieval mix and context caps (6 visual / 4 textual; max 4 images; 4000-char text) = 6:4 top-10; 4000 chars; 4 images
    Hand-set RAG context composition that shapes agent answers independent of learned models.
  • Species/centroid/SVM decision thresholds and top-N crop retention = configurable (not fixed globally)
    Configurable reject/fallback ranks and redundancy filters that gate what enters ChromaDB as taxonomic hypotheses.
  • Hypothetical daily block frequencies n_i / λ_event and duty interval Δt = scenario-dependent
    Energy totals E_total depend on assumed wake schedules and event rates rather than measured campaign statistics (§6.3.1).
  • QAT start epochs and train filters (≤5 boxes/image; 12 mammal species + noise) = QAT from epoch 73 (vision) / 37 (audio); ≤5 fish/image
    Training protocol choices that define reported sentinel metrics and may not match open-water class mix.
axioms (5)
  • domain assumption Continuous high-performance GPU sensing is incompatible with multi-month battery underwater deployments, so hierarchical sentinel→master activation is required.
    Stated throughout Introduction and §4.2 as design premise; not re-derived from first principles.
  • domain assumption Frozen BioCLIP-2/OpenCLIP embedding space plus centroids/linear SVMs yields usable hierarchical taxonomic hypotheses without per-mission backbone retraining.
    §4.3.3–4.3.4; cross-dataset species results show this is only weakly true under domain shift.
  • domain assumption MAX78000/02 CNN accelerator op-set and memory map constrain deployable sentinel models to supported conv/pool patterns and quantization.
    §5 hardware constraints; standard for this platform family.
  • ad hoc to paper Block-additive energy model E_total = E_sentinel + Σ n_i E_b_i with negligible idle gaps adequately predicts mission energy.
    §6.3.1 Eqs. 18–20; assumes measured blocks compose linearly under field wake patterns.
  • ad hoc to paper Sandboxed LLM-generated Pandas code and local instruct models (Qwen3.5-4B / Ornith-9B) are reliable enough for analyst and reporting paths on-device.
    §4.4.2 and §6.3 qualitative tests; authors note small VLMs remain unreliable for agentic tasks (§7.4).
invented entities (2)
  • Unified multi-agent control layer (Router, Analyst, Energy/Sensor, Dynamic Model Deployment, Reporting agents) no independent evidence
    purpose: Coordinate retrieval path selection, structured analytics, power policy, sentinel re-flashing, and Markdown scientific reports at the edge.
    Software architecture roles, not new physical objects; existence is the implementation itself.
  • Detection-aware local multimodal pipeline (Grounding DINO/YOLO-World crops → BioCLIP index → centroid/SVM ID → ChromaDB RAG) independent evidence
    purpose: Turn sentinel events into searchable taxonomic hypotheses and grounded answers without cloud.
    Composition of existing models/DB; novelty is integration and marine reference wiring.

pith-pipeline@v1.2.0-grok45-kimik3 · 35119 in / 3864 out tokens · 88310 ms · 2026-07-31T18:24:40.589284+00:00 · methodology

0 comments
read the original abstract

Marine life monitoring is limited by strict energy constraints, poor underwater connectivity, and the high cost of transmitting raw multimodal data from remote deployments. This paper proposes a low-consumption underwater monitoring architecture that combines always-on edge sensing with selective high-performance local reasoning. The system follows a hierarchical master--satellite design in which ultra-low-power MAX78000/MAX78002 microcontrollers continuously monitor visual and acoustic signals, while an NVIDIA Jetson Orin NX is activated only for scheduled processing, event-driven analysis, or researcher interaction. Once active, the Jetson executes a fully local multimodal pipeline for data ingestion, visual target extraction, embedding-based indexing, species identification, retrieval-augmented reasoning, and automated reporting. BioCLIP/OpenCLIP embeddings are used to organize mission data, marine taxonomic references, scientific documents, and operational metadata in local ChromaDB collections. A dedicated identification layer combines visual similarity search, centroid-based classification, and supervised classifiers to support adaptive species recognition. A LangChain-based multi-agent framework coordinates query routing, structured analysis, energy management, hardware reconfiguration, and report generation. The architecture is evaluated through visual and acoustic monitoring case studies. The proposed system bridges ultra-low-power continuous sensing with local multimodal intelligence, enabling underwater stations to produce structured, researcher-ready knowledge while compressing local data for flexible acoustic, optical, or satellite transmission, minimizing both energy use and communication overhead.

Figures

Figures reproduced from arXiv: 2607.24313 by Laurent Gautier, Mohamed Amine Janati, St\'ephane Barbot.

Figure 1
Figure 1. Figure 1: Acoustic data generation pipeline. This pipeline is strictly designed to remain memory efficient, because the same preprocessing steps must be executed locally on the MAX78000 in real time, while avoiding biases that could lead to misclassification [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Hierarchical, low-power underwater monitoring architecture. Edge-level microcontrollers (MAX78000/02) continuously monitor environmental events and selectively wake a high-performance master node (Jetson Orin NX) for local multimodal processing, retrieval-augmented generation, and researcher interaction [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Experimental implementation of the system This mode is suitable for deployments in which relevant biological or environmental events must be analyzed soon after detection, while continuous operation of the high￾power compute module remains undesirable. To implement this behavior, the event detection sets microcontrollers’ I/O that are connected to a shared TTL wake line, as illustrated in [PITH_FULL_IMAGE… view at source ↗
Figure 4
Figure 4. Figure 4: Hardware wake-up and power-hold circuit used in Event-Driven Hybrid Mode. power-switching stage to disconnect the 12 V supply. Con￾sequently, short detection pulses are sufficient to initiate a complete processing cycle without requiring the detecting microcontroller to remain active throughout the Jetson oper￾ating period. 4.2.3. Mission and Researcher-Interactive Mode Mission and Researcher-Interactive M… view at source ↗
Figure 6
Figure 6. Figure 6: Example of image-image similarity search using the Web UI [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example of text-image similarity search (iv) retrieval-augmented reasoning. This modular organiza￾tion clearly separates data management from species identi￾fication, making the role of each stage explicit and facilitat￾ing future extensions. Before the pipeline can be executed, the species identi￾fication layer must be initialized. During this configuration stage, a set of reference taxonomic datasets is … view at source ↗
Figure 8
Figure 8. Figure 8: Local multimodal processing and retrieval-augmented memory pipeline. embedding stage, as it relies on embedding representations introduced in the preceding section. 4.3.1. Detection-Aware Data Ingestion The pipeline begins with detections produced by the low￾power sentinel subsystems. These detections may include acoustic frames, seismic waves, visual observations or asso￾ciated metadata. In Autonomous Bat… view at source ↗
Figure 9
Figure 9. Figure 9: Unified agentic reasoning and reporting layer. Incoming researcher queries or automated system triggers are evaluated by a RouterAgent, which selects the appropriate retrieval, analysis, or reporting pathway. The agents interact with local memory, the local instruct model, and hardware-management modules to generate grounded outputs for researchers. station to ensure it survives the designated mission dura… view at source ↗
Figure 10
Figure 10. Figure 10: Visual detection subsystem. 5.1. Visual Detection Subsystem The proposed visual subsystem implements a sequential, shared-bus data architecture to optimize memory utilization during hardware-accelerated execution. Data acquisition be￾gins through both a MIPI CSI-2 and DVP interface, which captures RGB565 image frames and streams them sequen￾tially into external QSPI SRAM using a dedicated line handler. Th… view at source ↗
Figure 11
Figure 11. Figure 11: Acoustic detection subsystem. During CNN inference, the primary microcontroller yields execution to the hardware accelerator to reduce dy￾namic power consumption. Once inference is complete, the accelerator returns a vector of class scores, which is processed using a softmax function to obtain the confidence associated with each category. If the highest-confidence prediction corresponds to a cetacean voca… 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

34 extracted references · 8 canonical work pages · 3 internal anchors

  1. [1]

    FishDet-M: A Unified Large-Scale Benchmark for Robust Fish Detection and CLIP-Guided Model Selection in Diverse Aquatic Visual Domains

    Abujabal, M., Saoud, L.S., Hussain, I., 2025. Fishdet-m: A unified large-scalebenchmarkforrobustfishdetectionandclip-guidedmodel selectionindiverseaquaticvisualdomains. doi:10.48550/arXiv.2507. 17859,arXiv:2507.17859

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2507.17859 2025

  2. [2]

    Aquaticclip: A vision-language foundation modelforunderwatersceneanalysis

    Alawode, B., Ganapathi, I.I., Javed, S., Werghi, N., Bennamoun, M., Mahmood, A., 2025. Aquaticclip: A vision-language foundation modelforunderwatersceneanalysis. doi:10.48550/arXiv.2502.01785, arXiv:2502.01785

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.01785 2025

  3. [3]

    MAX78002: Artificial Intelligence Micro- controller with Low-Power Convolutional Neural Network Accel- erator

    Analog Devices, 2022a. MAX78002: Artificial Intelligence Micro- controller with Low-Power Convolutional Neural Network Accel- erator. Analog Devices. URL:https://www.analog.com/media/en/ technical-documentation/data-sheets/max78002.pdf. data sheet

  4. [4]

    MAX78002 Artificial Intelligence Micro- controller with Low-Power Convolutional Neural Network Acceler- ator Data Sheet

    Analog Devices, 2022b. MAX78002 Artificial Intelligence Micro- controller with Low-Power Convolutional Neural Network Acceler- ator Data Sheet. Technical Report. Analog Devices, Inc. URL: https://www.analog.com/en/products/max78002.html

  5. [5]

    OzFish Dataset - Machine Learning Dataset for Baited Remote Underwater Video Stations

    Australian Institute of Marine Science, 2020. OzFish Dataset - Machine Learning Dataset for Baited Remote Underwater Video Stations. Available online:https://github.com/open-AIMS/ozfish. doi:10.25845/5E28F062C5097. accessed: June 2026

    doi:10.25845/5e28f062c5097 2020

  6. [6]

    Real-time reporting of baleen whale passive acoustic detections from ocean gliders

    Baumgartner,M.F.,Fratantoni,D.M.,Hurst,T.P.,Brown,M.W.,Cole, T.V.N., Van Parijs, S.M., Johnson, M., 2013. Real-time reporting of baleen whale passive acoustic detections from ocean gliders. The JournaloftheAcousticalSocietyofAmerica134,1814–1823.doi:10. 1121/1.4816406

    2013

  7. [7]

    Yolo- world: Real-time open-vocabulary object detection

    Cheng,T.,Song,L.,Ge,Y.,Liu,W.,Wang,X.,Shan,Y.,2024. Yolo- world: Real-time open-vocabulary object detection. doi:10.48550/ arXiv.2401.17270,arXiv:2401.17270

    Pith/arXiv arXiv doi:10.48550/arxiv.2401.17270 2024

  8. [8]

    FishBase

    Froese, R., Pauly, D., 2024. FishBase. World Wide Web electronic publication. URL:https://www.fishbase.org. accessed: June 2026

    2024

  9. [9]

    CLIBD: Bridging vision and genomics forbiodiversitymonitoringatscale,in:ProceedingsoftheThirteenth InternationalConferenceonLearningRepresentations

    Gong, Z., Wang, A.T., Huo, X., Haurum, J.B., Lowe, S.C., Taylor, G.W., Chang, A.X., 2025. CLIBD: Bridging vision and genomics forbiodiversitymonitoringatscale,in:ProceedingsoftheThirteenth InternationalConferenceonLearningRepresentations. doi:10.48550/ arXiv.2405.17537

    work page doi:10.48550/arxiv.2405.17537 2025

  10. [10]

    Radio-triggered wake-up for wireless sensor networks

    Gu, L., Stankovic, J.A., 2005. Radio-triggered wake-up for wireless sensor networks. Real-Time Systems 29, 157–182. doi:10.1007/ s11241-005-6883-z

    2005

  11. [11]

    Irfan, M., Jiangbin, Z., Ali, S., Iqbal, M., Masood, Z., Hamid, U.,

  12. [12]

    Khan, F.F., Li, X., Temple, A.J., Elhoseiny, M., 2023. Fishnet: A large-scale dataset and benchmark for fish recognition, detection, and functional trait prediction, in: 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 20439–20449. doi:10. 1109/ICCV51070.2023.01874

    arXiv 2023

  13. [13]

    Syrene:Anunderwater embeddedartificialintelligencecameraforinvasivefaunamonitoring

    Laurent,G.,Barbot,S.,LeVourc’h,D.,2025. Syrene:Anunderwater embeddedartificialintelligencecameraforinvasivefaunamonitoring. Marine Technology Reporter Magazine 68, 40–42. URL:https: //archimer.ifremer.fr/doc/00992/110327/

    2025

  14. [14]

    doi:10.48550/arXiv.2007.10319, arXiv:2007.10319

    Lin,J.,Chen,W.M.,Lin,Y.,Cohn,J.,Gan,C.,Han,S.,2020.Mcunet: Tiny deep learning on iot devices. doi:10.48550/arXiv.2007.10319, arXiv:2007.10319

    Pith/arXiv arXiv doi:10.48550/arxiv.2007.10319 2020

  15. [15]

    Liu,S.,Zeng,Z.,Ren,T.,Li,F.,Zhang,H.,Yang,J.,Jiang,Q.,Li,C., Yang,J.,Su,H.,Zhu,J.,Zhang,L.,2024. Groundingdino:Marrying dino with grounded pre-training for open-set object detection, in: Computer Vision – ECCV 2024: 18th European Conference, Mi- lan, Italy, September 29–October 4, 2024, Proceedings, Part XLVII, Springer-Verlag, Berlin, Heidelberg. p. 38–5...

    doi:10.1007/978-3-031-72970-6_3 2024

  16. [16]

    Fish-vista: A multi-purpose dataset for understanding & identification of traits from images, in: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Mehrab, K.S., Maruf, M., Daw, A., Neog, A., Manogaran, H.B., Khurana, M., Feng, Z., Altintas, B., Bakis, Y., Campolongo, E.G., Thompson, M.J., Wang, X., Lapp, H., Berger-Wolf, T., Mabee, P., Bart, H., Chao, W.L., Dahdul, W.M., Karpatne, A., 2025. Fish-vista: A multi-purpose dataset for understanding & identification of traits from images, in: 2025 IEEE/CV...

    arXiv 2025

  17. [17]

    sleepy- cam: Power management mechanism for wireless video-surveillance cameras,in:2017IEEEInternationalConferenceonCommunications Workshops (ICC Workshops), pp

    Mekonnen, T., Harjula, E., Koskela, T., Ylianttila, M., 2017. sleepy- cam: Power management mechanism for wireless video-surveillance cameras,in:2017IEEEInternationalConferenceonCommunications Workshops (ICC Workshops), pp. 91–96. doi:10.1109/ICCW.2017. 7962639

    work page doi:10.1109/iccw.2017 2017

  18. [18]

    Millar, J., Huang, Y., Sethi, S., Haddadi, H., Madhavapeddy, A.,

  19. [20]

    Moosmann, J., Giordano, M., Vogt, C., Magno, M., 2023b. Tinyis- simoYOLO: A quantized, low-memory footprint, TinyML object detection network for low power microcontrollers, in: 2023 IEEE Mohamed Amine JANATI et al.:Preprint submitted to ElsevierPage 24 of 25 Energy Constrained Hierarchical Underwater Monitoring via Local Multi-Agent RAG 5th International ...

    arXiv 2023

  20. [21]

    Flexible and fully quantized lightweight TinyissimoYOLO for ultra-low-power edge systems

    Moosmann, J., Müller, H., Zimmerman, N., Rutishauser, G., Benini, L., Magno, M., 2024. Flexible and fully quantized lightweight TinyissimoYOLO for ultra-low-power edge systems. IEEE Access 12, 75093–75107. doi:10.1109/ACCESS.2024.3404878

    arXiv 2024

  21. [22]

    Fishnet: Deep neural networks for low-cost fish stock estimation, in: 2024 IEEE International Conference on Omni- layer Intelligent Systems (COINS), IEEE

    Mots’oehli, M., Nikolaev, A., IGede, W.B., Lynham, J., Mous, P.J., Sadowski, P., 2024. Fishnet: Deep neural networks for low-cost fish stock estimation, in: 2024 IEEE International Conference on Omni- layer Intelligent Systems (COINS), IEEE. pp. 1–7. doi:10.1109/ coins61597.2024.10622134

    arXiv 2024

  22. [23]

    Automated species classification and counting by deep-sea mobile crawler platforms using YOLO

    Ortenzi, L., Aguzzi, J., Costa, C., Marini, S., D’Agostino, D., Thom- sen, L., De Leo, F.C., Correa, P.V., Chatzievangelou, D., 2024. Automated species classification and counting by deep-sea mobile crawler platforms using YOLO. Ecological Informatics 82, 102788. doi:10.1016/j.ecoinf.2024.102788

    arXiv 2024

  23. [24]

    AntPi: A raspberry pi based edge-cloud system for real-time ant species detection using YOLO

    Palazzetti,L.,Giannetti,D.,Verolino,A.,Grasso,D.A.,Pinotti,C.M., Betti Sorbelli, F., 2025. AntPi: A raspberry pi based edge-cloud system for real-time ant species detection using YOLO. Ecological Informatics 91, 103383. doi:10.1016/j.ecoinf.2025.103383

    arXiv 2025

  24. [25]

    SeaLifeBase

    Palomares, M.L.D., Pauly, D., 2024. SeaLifeBase. World Wide Web electronicpublication. URL:https://www.sealifebase.org.accessed: June 2026

    2024

  25. [26]

    Learning transferable visual models from nat- ural language supervision, in: Proceedings of the 38th International Conference on Machine Learning, PMLR

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agar- wal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I., 2021. Learning transferable visual models from nat- ural language supervision, in: Proceedings of the 38th International Conference on Machine Learning, PMLR. pp. 8748–8763

    2021

  26. [27]

    A review on tinyml: State-of-the-art and prospects

    Ray, P.P., 2022. A review on tinyml: State-of-the-art and prospects. Journal of King Saud University - Computer and Information Sci- ences 34, 1595–1623. doi:10.1016/j.jksuci.2021.11.019

    work page doi:10.1016/j.jksuci.2021.11.019 2022

  27. [28]

    FMRAG: Retrieval-augmented multimodallargelanguagemodelsforfisheriesintelligence

    Song, Y., Lv, C., Zhu, K., 2026. FMRAG: Retrieval-augmented multimodallargelanguagemodelsforfisheriesintelligence. Frontiers in Marine Science 13, 1801835. doi:10.3389/fmars.2026.1801835

    arXiv 2026

  28. [29]

    BioCLIP: A Vision Foundation Model for the Tree of Life

    Stevens,S.,Wu,J.,Thompson,M.J.,Campolongo,E.G.,Song,C.H., Carlyn, D.E., Dong, L., Dahdul, W.M., Stewart, C., Berger-Wolf, T., Chao, W.L., Su, Y., 2024. BioCLIP: A vision foundation model for the tree of life, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). doi:10.48550/ arXiv.2311.18803

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2311.18803 2024

  29. [30]

    Reliable and efficient integration of AI into camera traps for smart wildlife monitoring based on continual learning

    Velasco-Montero, D., Fernández-Berni, J., Carmona-Galán, R., San- glas, A., Palomares, F., 2024. Reliable and efficient integration of AI into camera traps for smart wildlife monitoring based on continual learning. Ecological Informatics 83, 102815. doi:10.1016/j.ecoinf. 2024.102815

    arXiv 2024

  30. [31]

    SOUND Database of Marine Animal Vocalizations: Structure and Operations

    Watkins, W.A., Fristrup, K., Daher, M.A., Howald, T.J., 1992. SOUND Database of Marine Animal Vocalizations: Structure and Operations. Technical Report. Woods Hole Oceanographic Institu- tion. doi:10.1575/1912/854

    work page doi:10.1575/1912/854 1992

  31. [32]

    BioTrove: A Large Curated Image Dataset Enabling AI for Biodiversity

    Yang, C.H., Feuer, B., Jubery, Z., Deng, Z.K., Nakkab, A., Hasan, M.Z.,Chiranjeevi,S.,Marshall,K.,Baishnab,N.,Singh,A.K.,Singh, A., Sarkar, S., Merchant, N., Hegde, C., Ganapathysubramanian, B., 2024. Biotrove: A large curated image dataset enabling ai for biodiversity. arXiv preprint arXiv:2406.17720 doi:10.48550/arXiv. 2406.17720

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.17720 2024

  32. [33]

    Wildfish++:Acomprehensive fish benchmark for multimedia research

    Zhuang,P.,Wang,Y.,Qiao,Y.,2021. Wildfish++:Acomprehensive fish benchmark for multimedia research. IEEE Transactions on Multimedia 23, 3603–3617. doi:10.1109/TMM.2020.3028482. Mohamed Amine JANATI et al.:Preprint submitted to ElsevierPage 25 of 25

    arXiv 2021

  33. [2021]

    Expert Systems with Applications 183, 115270

    Deepship: An underwater acoustic benchmark dataset and a separable convolution based autoencoder for classification. Expert Systems with Applications 183, 115270. doi:10.1016/j.eswa.2021. 115270

    work page doi:10.1016/j.eswa.2021 2021

  34. [2025]

    doi:10.48550/arXiv

    Benchmarking ultra-low-power𝜇npus. doi:10.48550/arXiv. 2503.22567,arXiv:2503.22567

    doi:10.48550/arxiv.2503.22567