Pith. sign in

REVIEW 4 major objections 4 minor 3 cited by

A new 7,000-pair dataset links smell to sight and shows that vision can teach machines to recognize odors in the wild.

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 →

New York Smells is an in-the-wild dataset of 7,000 co-captured image–e-nose smell pairs covering 3,500 objects, and contrastive vision-smell training on it yields olfactory representations that outperform hand-crafted features.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Genuinely novel in-the-wild smell-vision dataset, but the raw-vs-smellprint claim needs a baseline ablation before I'd trust it. the 4 major comments →

arxiv 2511.20544 v2 pith:PFRAIOW4 submitted 2025-11-25 cs.CV cs.AIcs.LG

New York Smells: A Large Multimodal Dataset for Olfaction

classification cs.CV cs.AIcs.LG
keywords olfactionelectronic nosemultimodal learningcontrastive learningdatasetsmell-to-image retrievalrepresentation learningin-the-wild sensing
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 reading

This paper attempts to establish that co-located vision can serve as a teacher for olfaction. It introduces a dataset of 7,000 smell-image pairs from 3,500 objects recorded in natural indoor and outdoor settings with a handheld 32-sensor electronic nose and a camera. Using contrastive learning on the raw sensor signals, the authors show that the learned olfactory representations outperform the widely used hand-crafted smellprint feature across smell-to-image retrieval, scene/object/material recognition from smell alone, and fine-grained grass species discrimination. If the claim holds, machine olfaction can move beyond controlled laboratories into everyday environments, and training can rely on synchronized sight instead of costly chemical analyses.

Core claim

The paper's central claim is that paired visual and olfactory signals, captured together in the wild, enable cross-modal olfactory representation learning that beats hand-crafted features. Concretely, training a contrastive joint embedding between raw e-nose time series and synchronized images yields a smell encoder that substantially outperforms the standard smellprint descriptor on three benchmark tasks: retrieving the matching image from a smell query, recognizing scenes, objects, and materials from smell alone, and discriminating between two co-located grass species. The authors attribute this to the richer information present in the raw 28×32 sensor matrix compared to the 32-dimensional

What carries the argument

The central object is the raw olfactory signal matrix from the Cyranose 320 electronic nose: 10 seconds of ambient baseline followed by two 10-second samples of the target object, concatenated over 32 sensors into a 28×32 time-series. The argument is carried by a contrastive learning objective (termed COIP) that aligns this signal with synchronized images, learning a shared smell-sight embedding. The baseline comparator is the smellprint, a hand-crafted 32-dimensional feature computed as relative sensor response (sample peak minus baseline, divided by baseline) after Savitzky–Golay filtering.

Load-bearing premise

The 10-second sample stage, recorded after the ambient baseline, is assumed to reflect the target object's odor rather than the surrounding scene, and the co-located images are assumed to correspond to that same object; if ambient background dominates, the contrastive learning mostly aligns images with scene-level smell.

What would settle it

Query the trained model with olfactory recordings taken with the snout sealed or pointed at empty air in the same scenes; if retrieval accuracy remains comparable to the results with real object samples, the embedding is exploiting scene background rather than object odor. Alternatively, swapping the baseline segment for one from a different scene while keeping the sample segment should shift predictions if the object odor is the true signal.

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

If this is right

  • Smell-to-image retrieval becomes feasible in the wild: raw-signal encoders reach roughly 43% recall@20 compared to about 6% for the smellprint baseline.
  • Scene recognition from smell alone reaches around 99.5% accuracy with a CNN, showing that ambient olfactory context is strongly encoded in the raw signal.
  • Learned representations beat hand-crafted smellprints across all three benchmark tasks, including fine-grained discrimination between two grass species coexisting on the same lawn.
  • The dataset's scale—about 70 times more distinct objects than existing lab-collected olfaction datasets—opens the door to data-driven olfaction research outside controlled settings.
  • Visual supervision supplies a label-free training signal for olfaction, avoiding the need for costly perceptual descriptors or molecular analyses.

Where Pith is reading between the lines

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

  • If ambient scene odor dominates the 28×32 signal, the contrastive loss may align images with background smell rather than the probed object; a direct test would be to query with samples of empty air or to ablate the baseline stage.
  • The high scene-recognition accuracy could partly reflect this scene-level confound, so downstream users should report object classification with environmental variation held out.
  • The grass discrimination result is the strongest evidence for genuine object-level olfactory signal, since the two species are co-located; extending the benchmark to multiple co-located objects per scene would further validate the object-level claim.
  • A transfer test to a different e-nose or sensor array would clarify whether the learned embedding captures generic chemical properties or device-specific artifacts.
Share X Bluesky LinkedIn Reddit HN

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 / 4 minor

Summary. The paper introduces New York Smells, a claimed in-the-wild paired vision-olfaction dataset of 7,000 smell-image pairs from 3,500 objects, with 70x more objects than prior olfactory datasets. The authors mount a Cyranose 320 e-nose with a camera, record a 10-second ambient baseline followed by two 10-second snout samples per object, and concatenate them into a 28x32 raw signal. Object and material labels are generated by GPT-4o from the accompanying images. Using a contrastive learning objective (COIP, Eq. 1-2), they train smell and image encoders and evaluate on three tasks: smell-to-image retrieval, scene/object/material classification from smell, and grass-species discrimination. They report that raw-signal encoders outperform hand-crafted smellprint features across all tasks.

Significance. If the dataset and results hold, this is a substantial community resource: it is much larger and more naturalistic than existing e-nose datasets, it is the first to pair in-the-wild olfaction with images, and it demonstrates a plausible route to self-supervised olfactory representations. The authors commit to releasing data and code, and the dataset collection protocol is clearly described. However, the central scientific claim—that the learned representations capture object-level odor rather than scene/background odor—is not yet established because the raw signal includes the ambient baseline block and no ablation isolates the target-object contribution. The evaluation also lacks error bars, and the labels are machine-generated without verification. These issues are load-bearing for the claimed advances.

major comments (4)
  1. [Sec. 3.1, Fig. 5 and Sec. 5.2, Table 2] The raw olfactory signal is defined as the concatenation of a 10-second ambient baseline (purge inlet) and two 10-second snout samples, yielding a 28x32 matrix. The contrastive encoder thus has direct access to the ambient background block. The near-ceiling scene accuracy of the raw CNN (99.5% scratch) versus the smellprint (42.2%) is consistent with the hypothesis that the model aligns images with scene-level background odor rather than with object-emitted VOCs. This is a load-bearing confound for the retrieval and object/material results. Please add ablations: (a) train raw encoders on sample-minus-baseline, on baseline-only, and on sample-only inputs; (b) report retrieval and classification metrics split by whether the distractor/query share the same scene; and (c) quantify the signal-to-noise ratio between the sample and baseline windows for low-odor materials (e.g., metal, stone, pl
  2. [Sec. 3.1 'Labeling the dataset' and Sec. 5.2] Object and material labels are generated exclusively by GPT-4o from the visual stream, with no reported human verification or agreement measure. These labels define the held-out evaluation sets for Tables 2 and 3. If the VLM labels are noisy or biased (e.g., guessing 'plant shrub' from visual context rather than the probed object), the classification accuracies are not a valid measure of olfactory discriminability. Please report a human-verified subset (even a few hundred samples), per-category label reliability, or an inter-annotator agreement between GPT-4o and human raters. Also report how many samples were labeled 'unlabeled' and how these are handled in evaluation.
  3. [Tables 1-3 and Sec. 5.1] No error bars, confidence intervals, or significance tests are reported anywhere. The retrieval test set is N=933; the fine-grained grass task uses only 42 held-out samples. Differences between architectures (e.g., CNN vs Transformer retrieval recall @20: 32.6 vs 43.1) and between raw and smellprint may be within noise, especially given the small fine-grained test set. Please run multiple training seeds or bootstrap over test samples and report mean +/- std (or CIs). This is required to support the quantitative claims of superiority over hand-crafted features.
  4. [Sec. 5.1 retrieval protocol] The distractor sampling procedure is underspecified. The text says 'we sample a distractor set of images' but does not state whether distractors are drawn from the whole test set, whether same-scene images are excluded, or how N=933 is derived. If distractors include images from the same scene, retrieval can succeed by matching scene-level background, which would inflate recall. Specify the distractor distribution and, ideally, report retrieval after removing same-scene distractors and after ablating the baseline block.
minor comments (4)
  1. [Fig. 8 and Sec. 7.2] Typographical errors in labels: 'Planets Shrub' should be 'Plants Shrub', 'Treet Parts' should be 'Tree Parts'. The color-coding description in Sec. 7.2 says blue for objects and green for materials, but Fig. 8 caption may be inconsistent with the actual rendering; please check.
  2. [Sec. 4.2 and Eq. 5] The smellprint definition is clear, but the Savitzky-Golay filter parameters (window length w, polynomial order p) are never specified. Since the baseline comparison depends on these parameters, please provide the exact values used in all experiments.
  3. [Sec. 3.1] The dataset split description says 'uniformly split' but also requires both samples of an object to be in the same split. Clarify whether the split is by object or by scene/session, and report the number of distinct scenes and objects in train vs test.
  4. [References] Reference [15] is cited as 'concurrent, unpublished work' and appears twice in the related work and once in the introduction. If it has been published or updated, please use the final version and avoid repeating the same citation in adjacent sentences.

Circularity Check

0 steps flagged

No significant circularity: the paper's claims are empirical benchmark results with independent baselines; the ambient-baseline concern is a validity issue, not a definitional reduction.

full rationale

New York Smells is an empirical dataset and benchmark paper; it does not purport to derive a theoretical result from first principles. The central claims—dataset scale, retrieval/recognition accuracy, and raw-signal superiority over smellprint—are evaluated on held-out splits and against independently defined baselines. I checked each candidate pattern. (1) The VLM-generated object/material labels are derived from the same images that provide the COIP contrastive supervision, but the recognition protocol trains linear probes on training labels and evaluates on held-out test smells; the test label is not an input to the encoder, so accuracy is not forced by construction. (2) The smellprint Eqs. 3–5 is a hand-crafted relative-response feature defined independently of the learned encoders; the raw-vs-smellprint comparison, whatever its confounds, is an empirical comparison rather than a renaming of inputs. (3) The Sec. 3.1 protocol concatenates a 10-s ambient baseline with two 10-s snout samples, and a critic could argue the raw model's gains come from scene-level background odor encoded in the baseline block rather than object-level odor. That is a potential experimental confound and a robustness/validity concern, not a definitional equivalence: the prediction is still a function of the sensor signal, and the paper's equations do not make the target quantity equal to an input by construction. (4) No load-bearing self-citations or imported uniqueness theorems appear; the cited multimodal-learning works (e.g., CLIP, CMC) are standard external methods, and the concurrent SmellNet citation is a comparison baseline, not a premise. Under the hard rules, I therefore report no significant circularity. Score 0.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

No mathematical derivation is present; the central claims are empirical and rest on the sensor protocol, label quality, and dataset split design. The most important assumptions are the ambient-baseline independence and the accuracy of VLM-generated labels.

free parameters (3)
  • Contrastive temperature tau = 0.07
    Set by hand in Eq. 1; no tuning study or sensitivity analysis is reported.
  • Savitzky-Golay filter window and polynomial order (w, p)
    Needed to compute the smellprint (Eq. 3-5), but w and p are not reported; this affects all smellprint baselines.
  • Network architecture and training hyperparameters
    MLP/CNN/Transformer encoders are named but layer counts, widths, embedding dimensions, learning rates, epochs, and batch sizes are not specified, so the central comparison cannot be exactly reproduced.
axioms (4)
  • domain assumption Co-occurring image and e-nose capture correspond to the same target object/odor (natural synchronization between smell and sight during olfactory observation).
    Stated in Sec. 3.1; contrastive loss (Eq. 1) and retrieval ground truth assume this pairing is meaningful.
  • domain assumption The baseline (ambient) phase followed by the sample phase isolates target object odor; the raw T×32 concatenation of both phases is a valid input.
    See Fig. 5; if the ambient baseline dominates, scene recognition and retrieval results are confounded by background odor.
  • domain assumption GPT-4o automatically generated object/material labels are accurate enough to serve as ground truth for evaluation.
    Sec. 3.1 and Appendix 7.1; no human verification or inter-annotator agreement is reported.
  • ad hoc to paper The 49 object categories and the Matador material taxonomy are a valid closed set for the collected data.
    Sec. 3.1; the category lists are chosen by the authors and are not externally validated.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of New York Smells: A Large Multimodal Dataset for Olfaction." pith.science (2026). https://pith.science/paper/PFRAIOW4

@misc{pith2026251120544,
  author       = {Pith},
  title        = {Pith review of: New York Smells: A Large Multimodal Dataset for Olfaction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PFRAIOW4}},
  note         = {Machine review of arXiv:2511.20544}
}
Share X Bluesky LinkedIn Reddit HN
abstract

While olfaction is central to how animals perceive the world, this rich chemical sensory modality remains largely inaccessible to machines. One key bottleneck is the lack of diverse, multimodal olfactory training data collected in natural settings. We present New York Smells, a large dataset of paired image and olfactory signals captured ``in the wild.'' Our dataset contains 7,000 smell-image pairs from 3,500 distinct objects across indoor and outdoor environments, with approximately 70$\times$ more objects than existing olfactory datasets. Our benchmark has three tasks: cross-modal smell-to-image retrieval, recognizing scenes, objects, and materials from smell alone, and fine-grained discrimination between grass species. Through experiments on our dataset, we find that visual data enables cross-modal olfactory representation learning, and that our learned olfactory representations outperform widely-used hand-crafted features.

Figures

Figures reproduced from arXiv: 2511.20544 by Alexandra Horowitz, Andrew Owens, Carl Vondrick, Ege Ozguroglu, Junbang Liang, Mia Chiquier, Michael DeTienne, Ruoshi Liu, Wesley Wei Qian.

Figure 1
Figure 1. Figure 1: Multimodal olfaction in-the-wild. (a) We present New York Smells: a diverse, multimodal dataset of natural olfactory signals and paired visual data. We show one sequence of images and smell signals that we obtained in a public park (one scene of many in our dataset). We use this dataset for in-the-wild multimodal olfactory learning tasks that were not possible with previous datasets: (b) learning cross-mod… view at source ↗
Figure 2
Figure 2. Figure 2: The New York Smells dataset. We collect a diverse dataset of paired sight and olfaction by visiting many locations within New York City and recorded a variety of materials (top rows) and objects (bottom rows) in different scenes. We show a selection of the captured images here. All samples have a corresponding olfactory signal captured from the Cyranose electronic nose. This is in contrast to sight, sound,… view at source ↗
Figure 3
Figure 3. Figure 3: Odorant analysis. We show the distribution of objects and materials in our dataset. We use these labels to define smell understanding benchmarks. that are widely used in prior work. We see this dataset as a step toward in-the-wild, multi￾modal olfactory perception, as well as a step toward link￾ing sight with smell. While olfaction has traditionally been approached in constrained settings, such as quality … view at source ↗
Figure 4
Figure 4. Figure 4: Capturing paired sight and olfaction. We walk through a variety of real-world scenes and capture paired olfaction and visual signals using a camera mounted to an e-nose on a custom 3D-printed sensor rig. We point the e-nose’s snout at each object or substance of interest and record multiple images and smell signals from different orientations. We also capture a suite of other supplementary modalities: dept… view at source ↗
Figure 5
Figure 5. Figure 5: Olfactory signal: The raw smell signal is T × 32 di￾mensions where T is the capture time. The first part of capture is the baseline phase, where the ambient background smell is sensed. The second part is the sample phase, where the smell of the object of interest is sensed. This example shows the response for a flower. ing air through its snout and exposing the sampled com￾pounds to an array of 32 sensors.… view at source ↗
Figure 6
Figure 6. Figure 6: Contrastive olfactory-image learning. To demon￾strate the effectiveness of our dataset, we train general-purpose olfactory representations using contrastive learning. We train the model to align co-occurring visual and smell signals. The visual encoder processes RGB images, while the olfaction encoder pro￾cesses time-series sensor data from an e-nose. for each sample, assigning each data collection session… view at source ↗
Figure 7
Figure 7. Figure 7: Cross-modal retrieval qualitative results. We use our joint embeddings to match smell to images. Given a query smell, we find the images in the dataset that are the closest match in embedding space. Each row shows a reference smell query along with the top 5 image retrievals predicted by our model. The ground-truth smell-image pair is highlighted in green. Smell Encoder Mean Rank ↓ Median Rank ↓ Recall @ 5… view at source ↗
Figure 8
Figure 8. Figure 8: Recognizing scenes, objects, and materials from smell. We show the top 3 predictions from linear probing on the smell encoder. The predictions are from smell alone and the image is shown for visualization purposes only. Predictions are organized by color: red indicates scene classification, green indicates object classification and blue indicates material classification. Scenes Materials Objects Method Inp… view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. See & Sniff: Learning Visuo-Olfactory Representations

    cs.CV 2026-06 unverdicted novelty 7.0

    Introduces SmellNet-V synthetic visuo-olfactory dataset and See & Sniff self-supervised framework that learns aligned representations and produces smell saliency maps.

  2. What Images Cannot Say: Language-Guided Olfactory Representation Learning

    cs.CV 2026-07 conditional novelty 6.0

    SCENT uses VLM-generated scene descriptions as a semantic bridge to align electronic-nose signals with visual and textual embeddings, improving cross-modal smell retrieval and enabling object-context odor disentanglement.

  3. Chasing Ghosts: A Simulation-to-Real Olfactory Navigation Stack with Optional Vision Augmentation

    cs.RO 2026-02 unverdicted novelty 6.0

    A simulation-to-real navigation policy enables a quadrotor to locate an odor source using only basic olfaction sensors and optional vision, validated in indoor real-world flights.

Reference graph

Works this paper leans on

50 extracted references · 4 linked inside Pith · cited by 3 Pith papers

  1. [1]

    Look, listen and learn

    Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. InProceedings of the IEEE international conference on computer vision, 2017. 2

  2. [2]

    Elec- tronic nose for smart identification of roofing and paving grade asphalt.Transportation Research Procedia, 40:4–11,

    Federico Autelitano, Erika Garilli, and Felice Giuliani. Elec- tronic nose for smart identification of roofing and paving grade asphalt.Transportation Research Procedia, 40:4–11,

  3. [3]

    Sound- net: Learning sound representations from unlabeled video

    Yusuf Aytar, Carl V ondrick, and Antonio Torralba. Sound- net: Learning sound representations from unlabeled video. Advances in neural information processing systems, 29,

  4. [4]

    Identification of salmonella-inoculated beef using a portable electronic nose system.Journal of Rapid Methods & Automation in Microbiology, 13(2):71–95,

    S Balasubramanian, S Panigrahi, CM Logue, M Marchello, and JS Sherwood. Identification of salmonella-inoculated beef using a portable electronic nose system.Journal of Rapid Methods & Automation in Microbiology, 13(2):71–95,

  5. [5]

    Electronic nose and visible-near infrared spectroscopy in fruit and veg- etable monitoring.Reviews in Analytical Chemistry, 36(4): 20160016, 2017

    Roberto Beghi, Susanna Buratti, Valentina Giovenzana, Si- mona Benedetti, and Riccardo Guidetti. Electronic nose and visible-near infrared spectroscopy in fruit and veg- etable monitoring.Reviews in Analytical Chemistry, 36(4): 20160016, 2017. 3

  6. [6]

    Learning to taste: A multimodal wine dataset.Advances in Neural Information Processing Systems, 36:7351–7360, 2023

    Thoranna Bender, Simon Sørensen, Alireza Kashani, Krist- jan Eldjarn Hjorleifsson, Grethe Hyldig, Søren Hauberg, Serge Belongie, and Frederik Warburg. Learning to taste: A multimodal wine dataset.Advances in Neural Information Processing Systems, 36:7351–7360, 2023. 3

  7. [7]

    Beveridge and S

    M. Beveridge and S. K. Nayar. Hierarchical Material Recog- nition from Local Appearance. InProc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025. 5

  8. [8]

    Vggsound: A large-scale audio-visual dataset

    Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zis- serman. Vggsound: A large-scale audio-visual dataset. In International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721–725. IEEE, 2020. 2

  9. [9]

    Craven, Eric G

    Brent A. Craven, Eric G. Paterson, and Gary S. Settles. The fluid dynamics of canine olfaction: unique nasal airflow pat- terns as an explanation of macrosmia.Journal of the Royal Society Interface, 7(47):933–943, 2010. 4

  10. [10]

    Learning classification with unlabeled data

    Virginia De Sa. Learning classification with unlabeled data. Advances in neural information processing systems, 6, 1993. 3

  11. [11]

    Predicting human odor perception represented by continuous values from mass spectra of essential oils resembling chemical mixtures.PLOS ONE, 15(6):e0234688, 2020

    Tanoy Debnath and Takamichi Nakamoto. Predicting human odor perception represented by continuous values from mass spectra of essential oils resembling chemical mixtures.PLOS ONE, 15(6):e0234688, 2020. 3

  12. [12]

    Assessment of the portable c-320 electronic nose for discrimination of nine insectivorous bat species: implica- tions for monitoring white-nose syndrome.Biosensors, 10 (2):12, 2020

    Anna C Doty, A Dan Wilson, Lisa B Forse, and Thomas S Risch. Assessment of the portable c-320 electronic nose for discrimination of nine insectivorous bat species: implica- tions for monitoring white-nose syndrome.Biosensors, 10 (2):12, 2020. 3

  13. [13]

    Tactile-augmented radiance fields.arXiv preprint arXiv:2405.04534, 2024

    Yiming Dou, Fengyu Yang, Yi Liu, Antonio Loquercio, and Andrew Owens. Tactile-augmented radiance fields.arXiv preprint arXiv:2405.04534, 2024. 3

  14. [14]

    Bacteria classification using cyranose 320 electronic nose.Biomedical engineering online, 1(1):4, 2002

    Ritaban Dutta, Evor L Hines, Julian W Gardner, and Pascal Boilot. Bacteria classification using cyranose 320 electronic nose.Biomedical engineering online, 1(1):4, 2002. 3, 7

  15. [16]

    Smell- net: A large-scale dataset for real-world smell recognition

    Dewei Feng, Carol Li, Wei Dai, and Paul Pu Liang. Smell- net: A large-scale dataset for real-world smell recognition. arXiv preprint arXiv:2506.00239, 2025. 4

  16. [17]

    First steps in developing a fast, cheap, and reliable method to distinguish wild mush- room and truffle species.Resources, 12(12):139, 2023

    In ˆes Ferreira, Teresa Dias, Juliana Melo, Abdul Mounem Mouazen, and Cristina Cruz. First steps in developing a fast, cheap, and reliable method to distinguish wild mush- room and truffle species.Resources, 12(12):139, 2023. 3, 7

  17. [18]

    Ad- vances in electronic nose sensors for plant disease and pest detection.Engineering Proceedings, 48(1):14, 2023

    Ana Fundurulic, Jorge MS Faria, and Maria L In ´acio. Ad- vances in electronic nose sensors for plant disease and pest detection.Engineering Proceedings, 48(1):14, 2023. 3

  18. [19]

    Canine olfaction: Scent, sign, and situation

    Simon Gadbois and Catherine Reeve. Canine olfaction: Scent, sign, and situation. InDomestic Dog Cognition and Behavior: The Scientific Study of Canis familiaris, pages 3–

  19. [20]

    Assessment of an e-nose performance for the detection of covid-19 specific biomark- ers.Journal of Breath Research, 17(2):026006, 2023

    Christelle Ghazaly, Krystyna Biletska, Etienne A Thevenot, Philippe Devillier, Emmanuel Naline, Stanislas Grassin- Delyle, and Emmanuel Scorsone. Assessment of an e-nose performance for the detection of covid-19 specific biomark- ers.Journal of Breath Research, 17(2):026006, 2023. 3

  20. [21]

    Imagebind: One embedding space to bind them all

    Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 15180–15190, 2023. 2

  21. [22]

    Hepper and D

    P. Hepper and D. Wells. Olfaction in the order carnivora: Family canidae. InHandbook of Olfaction and Gustation, pages 591–603. Wiley-Blackwell, Hoboken, NJ, 3 edition,

  22. [23]

    P. G. Hepper. The discrimination of human odor by the dog. Perception, 17:549–554, 1988. 4 9

  23. [24]

    Gerkin, Yinzhao Guan, Avi Dhu- randhar, Nath Tsaernstad, Jun Tan, Moustapha Bensafi, et al

    Andreas Keller, Richard C. Gerkin, Yinzhao Guan, Avi Dhu- randhar, Nath Tsaernstad, Jun Tan, Moustapha Bensafi, et al. Predicting human olfactory perception from chemical fea- tures of odor molecules.Science, 355(6327):820–826, 2017. 3

  24. [25]

    Canine olfaction: physiology, behavior, and pos- sibilities for practical applications.Animals, 11(8):2463,

    Agata Kokcinska-Kusiak, Martyna Woszczylo, Mikolaj Zy- bala, Julia Maciocha, Katarzyna Barlowska, and Michal Dzieciola. Canine olfaction: physiology, behavior, and pos- sibilities for practical applications.Animals, 11(8):2463,

  25. [26]

    Lee, Emily J

    Brian K. Lee, Emily J. Mayhew, Benjamin Sanchez- Lengeling, Jennifer N. Wei, Wesley W. Qian, Kelsie A. Lit- tle, Matthew Andres, Britney B. Nguyen, Theresa Moloy, Jacob Yasonik, Jane K. Parker, Richard C. Gerkin, Joel D. Mainland, and Alexander B. Wiltschko. A principal odor map unifies diverse tasks in olfactory perception.Science, 381(6661):999–1006, 2023. 3, 4

  26. [27]

    Changying Li, Paul Heinemann, and Richard Sherry. Neural network and bayesian network fusion models to fuse elec- tronic nose and surface acoustic wave sensor data for apple defect detection.Sensors and Actuators B: Chemical, 125 (1):301–310, 2007. 3

  27. [28]

    Scent classification by k nearest neigh- bors using ion-mobility spectrometry measurements.Expert Systems with Applications, 115:593–606, 2019

    Philipp M ¨uller, Katri Salminen, Ville Nieminen, Anton Kon- tunen, Markus Karjalainen, Poika Isokoski, Jussi Rantala, Mariaana Savia, Jari V ¨aliaho, Pasi Kallio, Jukka Lekkala, and Veikko Surakka. Scent classification by k nearest neigh- bors using ion-mobility spectrometry measurements.Expert Systems with Applications, 115:593–606, 2019. 3

  28. [29]

    Springer, Berlin, Heidelberg, 2014. 4

  29. [30]

    Multimodal deep learn- ing

    Jiquan Ngiam, Aditya Khosla, Mingyu Kim, Juhan Nam, Honglak Lee, Andrew Y Ng, et al. Multimodal deep learn- ing. InICML, pages 689–696, 2011. 3

  30. [31]

    Orthog- onal gas sensor arrays with intelligent algorithms for early warning of electrical fires.Sensors and Actuators B: Chemi- cal, 130(2):889–899, 2008

    Ming Ni, Joseph R Stetter, and William J Buttner. Orthog- onal gas sensor arrays with intelligent algorithms for early warning of electrical fires.Sensors and Actuators B: Chemi- cal, 130(2):889–899, 2008. 3

  31. [32]

    Visually indicated sounds

    Andrew Owens, Phillip Isola, Josh McDermott, Antonio Tor- ralba, Edward H Adelson, and William T Freeman. Visually indicated sounds. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2405–2413,

  32. [33]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2, 8

  33. [34]

    A measure of smell enables the creation of olfactory metamers.Nature, 588(7836):118–123, 2020

    Aharon Ravia, Kobi Snitz, Danielle Honigstein, Maya Finkel, Rotem Zirler, Ofer Perl, Lavi Secundo, Christophe Laudamiel, David Harel, and Noam Sobel. A measure of smell enables the creation of olfactory metamers.Nature, 588(7836):118–123, 2020. 3

  34. [35]

    Machine learning for scent: Learning general- izable perceptual representations of small molecules.arXiv preprint arXiv:1910.10685, 2019

    Benjamin Sanchez-Lengeling, Jennifer N Wei, Brian K Lee, Richard C Gerkin, Al ´an Aspuru-Guzik, and Alexander B Wiltschko. Machine learning for scent: Learning general- izable perceptual representations of small molecules.arXiv preprint arXiv:1910.10685, 2019. 3

  35. [36]

    Cyranose 320 electronic nose, 2000

    Sensigent. Cyranose 320 electronic nose, 2000. Product page. 3, 4

  36. [37]

    The impact of storage conditions on stool smellprints as assessed by an electronic nose.ACS sensors, 10(2):689–698, 2025

    W Christopher Shelley, Anthony R Pecoraro, Misty Good, Fikir M Mesfin, Krishna Manohar, John P Brokaw, An- gela M Hansen, Robert H Pepin, Jonathan A Karty, Troy B Hawkins, et al. The impact of storage conditions on stool smellprints as assessed by an electronic nose.ACS sensors, 10(2):689–698, 2025. 3, 7

  37. [38]

    Khan, and Noam Sobel

    Kobi Snitz, Ayelet Yablonka, Tal Weiss, Ilana Frumin, Re- han M. Khan, and Noam Sobel. Predicting odor perceptual similarity from odor structure.PLOS Computational Biol- ogy, 9(10):e1003184, 2013. 3

  38. [39]

    Con- trastive multiview coding, 2020

    Yonglong Tian, Dilip Krishnan, and Phillip Isola. Con- trastive multiview coding, 2020. 2, 8

  39. [40]

    Mixed explosives dataset, 2019

    Julio Torres-Tello, Ana Guaman, and Seok-Bum Ko. Mixed explosives dataset, 2019. 3

  40. [41]

    Repre- sentation learning with contrastive predictive coding, 2019

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Repre- sentation learning with contrastive predictive coding, 2019. 6

  41. [42]

    Smell—adding a new dimension to urinalysis.Biosensors, 10(5):48, 2020

    Eva H Visser, Daan JC Berkhout, Jiwanjot Singh, Anne- mieke Vermeulen, Niloufar Ashtiani, Nanne K de Boer, Joanna AE van Wijk, Tim G de Meij, and Arend B¨okenkamp. Smell—adding a new dimension to urinalysis.Biosensors, 10(5):48, 2020. 3, 7

  42. [43]

    O’Connell, and Elizabeth J

    Matt Wachowiak, Adam Dewan, Thomas Bozza, T. O’Connell, and Elizabeth J. Hong. Recalibrating olfactory neuroscience to the range of naturally occurring odor con- centrations.Journal of Neuroscience, 45(10):e1872242024,

  43. [44]

    Recalibrating olfactory neuroscience to the range of naturally occurring odor con- centrations.Journal of Neuroscience, 45(10), 2025

    Matt Wachowiak, Adam Dewan, Thomas Bozza, Tom F O’Connell, and Elizabeth J Hong. Recalibrating olfactory neuroscience to the range of naturally occurring odor con- centrations.Journal of Neuroscience, 45(10), 2025. 4

  44. [45]

    the sense of smell in our languages

    Leo Weisgerber. Das geruchssinn in unseren sprachen (“the sense of smell in our languages”), 1928. 6

  45. [46]

    Smell terms are not rara: A semantic investigation of odor vocabu- lary in thai.Linguistics, 58(4):937–966, 2020

    Ewelina Wnuk, Rujiwan Laophairoj, and Asifa Majid. Smell terms are not rara: A semantic investigation of odor vocabu- lary in thai.Linguistics, 58(4):937–966, 2020. 6

  46. [47]

    Touch and go: Learn- ing from human-collected vision and touch.arXiv preprint arXiv:2211.12498, 2022

    Fengyu Yang, Chenyang Ma, Jiacheng Zhang, Jing Zhu, Wenzhen Yuan, and Andrew Owens. Touch and go: Learn- ing from human-collected vision and touch.arXiv preprint arXiv:2211.12498, 2022. 2, 3

  47. [48]

    An odor is not worth a thousand words: from multidimensional odors to unidimen- sional odor objects.Annual review of psychology, 61(1):219– 241, 2010

    Yaara Yeshurun and Noam Sobel. An odor is not worth a thousand words: from multidimensional odors to unidimen- sional odor objects.Annual review of psychology, 61(1):219– 241, 2010. 6

  48. [49]

    Connecting look and feel: Associating the visual and tactile properties of physical materials

    Wenzhen Yuan, Shaoxiong Wang, Siyuan Dong, and Edward Adelson. Connecting look and feel: Associating the visual and tactile properties of physical materials. InProceedings of the IEEE conference on computer vision and pattern recog- nition, pages 5580–5588, 2017. 3

  49. [50]

    Feasibility of detection of infested rice using an electronic nose.Journal of Stored Products Research, 92:101805, 2021

    Molin Zhou, Ragab Khir, Zhongli Pan, James F Campbell, Randall Mutters, and Zhuoyan Hu. Feasibility of detection of infested rice using an electronic nose.Journal of Stored Products Research, 92:101805, 2021. 3 10

  50. [51]

    You are shown four images where a blue sensor probe with a yellow tip is pointing at the same object

    Appendix 7.1. VLM Prompt for Labeling The following Python function is used to label objects using GPT-4o, where images are passed to GPT-4o along with a structured prompt to select the closest matching object category. Listing 1. Object labeling with GPT-4o. 1def label_gpt_views(image_path1, image_path2, image_path3, image_path4, indexed_labels, labels):...

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.