{"id":"0fd0e512-052f-4fb9-8ab0-9508835c878a","arxiv_id":"2507.20757","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A speckle-vibrometry rig with a transformer can remotely classify the liquid level inside opaque containers from surface vibrations excited by sound.","lead":"This paper shows a camera-laser system that reads the hidden fill level of opaque bottles and cans by watching tiny vibrations on their surfaces, then classifies each container type and how full it is. A smart generalist might care because it does this remotely, without touching or weighing the contents, and could be used for warehouse or hazmat inspection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'invariant to the vibration source' claim rests on an untested non-flat-spectrum assumption: Eq. (1) keeps |H||X|, so narrowband or notch-heavy excitation can erase the resonances the model needs.","rationale":"The reader's weakest assumption is source invariance, and my stress-test identifies the same load-bearing concern. The paper's own Eq. (1) and footnote indicate the condition, but the abstract states invariance unconditionally and the validation set does not stress non-flat source spectra. The strong within-distribution results and the ambient-sound experiment are real evidence, and the smooth-filter augmentation may confer some robustness; the issue is scope, not internal inconsistency or bad faith. The concern is load-bearing because it targets the abstract's headline claim and the practical deployment scenario, but it is empirically testable and is not contradicted by the reported data. Since the reader already returned CONDITIONAL, I would not move the verdict; the appropriate action is to keep it conditional pending the excitation-spectrum stress test and the release of data or code for exact reproduction.","tokens_in":14044,"tokens_out":5894,"duration_ms":73516,"concrete_test":"Select a held-out set of containers and levels. Synthesize five excitation signals with controlled spectra: (i) a 100-2500 Hz chirp; (ii) the same chirp with a 30 Hz-wide notch placed at a known resonant peak of the container; (iii) narrowband 100/120 Hz harmonic hum; (iv) pink noise; (v) music with strong spectral nulls. Run the existing pretrained model on each excitation without fine-tuning. If MAE for (ii)-(v) rises substantially above the reported 0.04-0.16, source invariance is disproven. Then retrain on the same data with input |F{v_i}|/|F{x}|, using a recorded speaker reference or accelerometer signal, and repeat; if the normalized model is robust across all five sources, the paper should add excitation normalization and soften the invariance claim accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 defines the model input as V_i[f] = |F{v_i}| (Eq. 1), with footnote 1 adding that source invariance holds 'as long as the excitation signal is sufficiently broadband.' For a linear structure, V_i(f) = H_i(f)X(f), so the input is the product of the container transfer function and the excitation spectrum. If X(f) has deep notches or narrowband energy, resonances in H_i are masked or spurious peaks appear; no Fourier-magnitude input can separate the two without a reference measurement of X. The paper trains on a two-second chirp and a song segment, augments with smooth random filters, and evaluates with supermarket ambient noise (test d). These are broadband, positive-spectrum signals; no evaluation uses a source with strong spectral notches or a narrowband hum (e.g., 50/120 Hz harmonics), which are common in realistic deployment. The abstract's unconditional invariance claim is therefore stronger than what Eq. (1) and the experiments establish. This is load-bearing because the practical value of the method is passive, remote sensing in arbitrary acoustic environments: if a real ambient source colors the spectrum, the fill-level estimate can degrade substantially, and the reported MAE of 0.04-0.16 does not bound that failure mode.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a speckle-based vibrometry system that captures vibrations on a 2D grid of laser points using a single defocused camera, enabling simultaneous multi-point measurement of multiple containers. The authors introduce a transformer-based architecture (the Vibration Transformer) that takes the Fourier magnitudes of vibration signals at three surface points per container and classifies container type and discrete liquid fill level. They collect a dataset of everyday containers and report MAE values of 0.01 (within-distribution), 0.09 (unseen instances), and 0.04 (ambient sound), with ablations showing that a CNN baseline fails at level prediction and that multi-point measurements help in harder generalization settings. The paper claims invariance to the vibration source, qualified in a footnote as holding for 'sufficiently broadband' excitation.","tokens_in":14275,"tokens_out":7031,"duration_ms":82758,"significance":"If the results hold, this is a novel non-contact sensing modality for inspecting sealed containers, with potential applications in warehouses, industrial monitoring, and hazardous-liquid storage. The paper's strengths include a genuine hardware contribution (2D-grid speckle vibrometry with ROI-based high-speed readout), a newly collected dataset, a physics-inspired architecture with sensible ablations, and a self-critical limitations section. The experimental support is appropriate for a proof of concept. The main weakness is that the source-invariance claim is broader than the evidence supports, since only a limited set of broadband excitations were tested and the conditioning assumption 'sufficiently broadband' is not quantitatively defined.","major_comments":[{"comment":"The claim that the architecture is 'invariant to the vibration source' is load-bearing and currently rests on an untested conditioning assumption. As Eq. (1) shows, the model input is |F{v_i}|, which for a linear structure equals |H_i(f)||X(f)|; separating the container transfer function from the excitation spectrum is impossible without a reference measurement of X(f) when X(f) has deep notches or is narrowband. Footnote 1 limits the claim to 'sufficiently broadband' excitation, but this condition is never defined quantitatively and no experiment uses a source with strong spectral notches or a narrowband hum (e.g., 50/120 Hz harmonics), which are common in real environments. The three tested excitations (chirp, song, supermarket noise) and the random smooth-filter augmentation all have broad, positive spectra. Please either (i) add experiments with held-out broadband sources (e.g., white noise, a different song) and with narrowband or notched sources, reporting MAE for each, and discuss what 'sufficiently broadband' means quantitatively, or (ii) remove the unconditional invariance wording from the abstract and Introduction and state the limitation explicitly in Sec. 7.","section":"Abstract; §1, §4, §6(d–f)"}],"minor_comments":[{"comment":"The frequency set F_fixed = {100, 100.5, ..., 2500 Hz} contains 4801 values, but the text says the resulting V_i is a 2×4800 matrix; please correct the count or the frequency set.","section":"§5.2"},{"comment":"Excluding one of six speakers yields about 16.7% of the data for testing, not 'about 20%' as stated; please adjust the text.","section":"§6, test (a)"},{"comment":"The phrase 'Our GPU implementation is ×20 faster' lacks a baseline; please specify the comparison (e.g., a CPU implementation, or a previous version of PCLK) and the hardware.","section":"§3"},{"comment":"The abstract states 'invariant to the vibration source' without the 'sufficiently broadband' qualification from Footnote 1; adding the qualifier would align the abstract with the technical content of the paper.","section":"Abstract; §1"},{"comment":"The 'chance≈ 30%' figure for level MAE is not defined; a brief explanation (e.g., expected MAE of a uniform random guess on [0,1]) would improve clarity.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid proof-of-concept for a novel sensing modality, and the dataset and hardware contributions are valuable. The main risk is the overstatement of source invariance: the conditioning assumption in the footnote is not tested with narrowband or notched excitations, so the authors should either provide those experiments or temper the claim. The self-citations are appropriate and not a concealment of prior work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a solid proof-of-concept for a genuinely new task: inferring liquid levels inside opaque containers by measuring surface vibrations with a speckle-based imaging system. The engineering is the real contribution. The 2D grid illumination, defocused camera, and batched GPU phase-correlation/Lucas-Kanade pipeline let the authors capture three vibration points on each of six containers simultaneously at 5 kHz. That is a practical step forward from prior single-point or scanning vibrometry. The transformer architecture is not exotic, but it is sensibly matched to the frequency-domain input, and the experiments are more thorough than the average vision paper: held-out speaker positions, unseen same-class instances, unseen liquid levels, and ambient sound are all tested separately, with an honest ablation showing multi-point data matters for the hard cases and a CNN baseline that fails as expected. The math is simple and correct, and there is no circularity; the result is empirical through and through.\n\nThe main soft spot is the claim in the abstract that the architecture is \"invariant to the vibration source.\" That is too strong. Equation (1) is the Fourier magnitude of the measured vibration, which is |H(f)X(f)| for a linear system - the container transfer function times the excitation spectrum. You only recover |H| if X is sufficiently flat over the band of interest, which is exactly what footnote 1 says, but the abstract and parts of the introduction drop that qualification. The tests only use a chirp, a song segment, and supermarket ambient noise - all broadband, positive-spectrum signals. Nothing evaluates a narrowband hum or a source with deep spectral notches, which are common in real industrial environments. The smooth random filter augmentation is a nice attempt, but it is not the same as testing a real narrowband source. This is load-bearing for the stated application, so it needs fixing, either by softening the claim or by adding experiments with narrowband excitation and, ideally, a reference measurement of X.\n\nThe other issues are minor in comparison: no code or data release (reproduction is impossible), a small dataset (5910 samples), and a somewhat hand-wavy use of the expectation estimator for continuous levels. None of these undercut the core idea.\n\nWho should read this? Anyone working in non-contact sensing, vibrometry, or hidden property inference. It deserves a serious referee, not a desk rejection. I would send it to review and ask the authors to address the source-invariance claim directly - either add narrowband tests or rewrite the abstract to say \"broadband sources.\" Also push for releasing the dataset, which could be valuable to the community.","headline":"A genuinely new proof-of-concept for remote fill-level sensing via speckle vibrometry, with a real overclaim about source invariance that should be fixed before publication.","tokens_in":805,"tokens_out":1082,"would_cite":true,"duration_ms":34811,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a camera capturing laser speckle vibrations on a 2D grid of surface points, combined with a transformer, can remotely infer the fill level of opaque liquid containers to about 1% error within distribution, without…","keywords":["speckle vibrometry","liquid level estimation","opaque containers","non-contact sensing","vibration transformer","laser speckle","computer vision","object inspection"],"falsifier":"Play a single pure tone near a container's resonant frequency as the only excitation, and evaluate the trained model on an unseen instance of that container: if the source-invariance claim is correct, the MAE should stay close to the 0.04 seen with ambient sound, while a large MAE increase would show that narrowband or notched excitations break the Fourier-magnitude proxy.","tokens_in":13797,"feed_emoji":"🥤","tokens_out":6136,"duration_ms":65171,"temperature":0.7,"pith_summary":"This paper tries to give computer vision the ability to see inside opaque liquid containers. Its central claim is that a container's tiny surface vibrations, captured remotely by imaging laser speckle at many points at once, encode its hidden liquid level accurately enough to be learned by a transformer. The authors build a first-of-a-kind 2D-grid speckle vibrometry system, record a dataset of everyday containers, and train a Vibration Transformer that reads fill level from Fourier magnitudes of the vibrations. They report errors as low as 1% for containers seen during training, 9% on unseen instances of the same class, and 4% under ambient sound. If correct, the method opens a non-contact way to inspect sealed containers in warehouses, factories, and settings where weighing or touching is impractical or unsafe.","feed_headline":"Surface vibrations reveal hidden liquid levels in sealed containers","feed_subtitle":"Laser-speckle camera plus transformer measures fill without touching or opening the container.","key_machinery":"The mechanism is speckle-based vibrometry on a 2D grid: one laser is split into a 6x6 grid of points, projected onto the containers, and a defocused camera records each point as a patch of random interference (speckle) that shifts with surface tilt. The authors track these shifts with phase correlation followed by Lucas-Kanade (PCLK+) to get two-axis vibration signals per point at rates up to 57 kHz. The Vibration Transformer then takes the Fourier magnitudes of these signals over 100-2500 Hz, tokenizes frequency bands, processes each point with a PointTransformer, and fuses the points with a ShapeTransformer, trained with a SORD loss that respects the ordinal nature of fill levels.","core_discovery":"The paper's central discovery is that a 2D grid of remotely sensed surface vibrations, converted to per-point Fourier magnitudes, is enough to regress an opaque container's liquid level across many everyday container types. On the sensing side, the authors build a laser-grid speckle system that captures vibrations of several containers simultaneously at up to 57 kHz, and on the inference side a 'Vibration Transformer' whose shared PointTransformer encodes each point's spectrum and whose ShapeTransformer fuses the points to produce container class and fill level. They report a 0.01 MAE (1%) for within-distribution fill levels, 0.09 for unseen instances of a known class, and 0.04 under ambient sound, and they show the model can interpolate to fill levels (25%, 50%, 75%) it never trained on.","pith_inferences":["The source-invariance argument suggests a practical recipe: in industrial settings one could deliberately play a short broadband probe (a chirp or noise burst) and then rely on ambient sound afterward; the paper does not test this mixed-excitation scenario.","Since multi-point data clearly helps on unseen instances, extending the 3 points per container to a denser grid (the hardware already captures a 6x6 grid) could be expected to further improve generalization beyond the paper's reported results.","The same Fourier-magnitude representation would presumably apply to other hidden contents whose presence changes acoustics, such as granular materials, powders, or spoilage gases in sealed food; the authors name these as open questions, not demonstrated claims.","A narrowband ambient environment (e.g., a room with a strong 50/60 Hz hum or a tonal alarm) could violate the broadband assumption; a robust system might need to estimate and whiten the excitation spectrum, which the current model does not do."],"forward_implications":["Warehouse and factory inspection of sealed beverage or chemical containers could be done remotely and at once, with no physical handling or weighing.","Because the model discards phase and uses only Fourier magnitudes, the same trained network generalizes to different excitation sounds, including unseen ambient noise, as long as the sound is broadband.","The ordinal loss lets the model interpolate to fill levels never seen in training (e.g., 25%, 50%, 75%), so the method is not limited to the six discrete training levels.","Trained on several instances of a container class, the model can predict the fill level of a new instance of the same class, such as the sixth can of a six-pack."],"supporting_citations":[{"why":"Establishes that speaker-excited surface vibrations captured by camera can be used to estimate object properties; the methodological starting point for inferring liquid level from vibration.","marker":"[13]"},{"why":"The prior dual-shutter speckle vibration system that sensed a single row of points; this work extends the approach to a simultaneous 2D grid of points across multiple containers.","marker":"[46]"},{"why":"Demonstrates speckle-pattern shift as a measure of surface vibration from a defocused laser spot, the underlying sensing principle.","marker":"[57]"},{"why":"Phase correlation method used to estimate integer-pixel speckle shifts between consecutive frames.","marker":"[26]"},{"why":"Lucas-Kanade optical flow estimation used for the sub-pixel residual shifts after phase correlation.","marker":"[31]"},{"why":"Transformer architecture on which the Vibration Transformer's attention-based design is based.","marker":"[52]"},{"why":"Introduces soft labels for ordinal regression, the basis of the SORD loss used for liquid-level classification.","marker":"[16]"},{"why":"Prior approach that infers liquid level from sound resonance after physically knocking on a container; the baseline the paper's remote visual method is designed to surpass.","marker":"[21]"}],"fun_headline_variants":["Laser speckle vibrations expose liquid levels in sealed containers","Vibration sensing reads fill levels without opening the container","Remote vibration analysis reveals hidden liquid amounts","Speckle vibrometry measures liquid levels through opaque walls","Transformers decode vibrations to gauge liquid contents"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Fourier magnitude of the recorded vibrations is a faithful proxy for the container's transfer function, which holds only when the excitation sound is sufficiently broadband and which the paper tests only with a chirp, a song segment, and one ambient noise type.","fun_headline_variants_meta":{"raw":{"variants":["Laser speckle vibrations expose liquid levels in sealed containers","Vibration sensing reads fill levels without opening the container","Remote vibration analysis reveals hidden liquid amounts","Speckle vibrometry measures liquid levels through opaque walls","Transformers decode vibrations to gauge liquid contents"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000483,"raw_usage":{"total_tokens":2394,"prompt_tokens":959,"completion_tokens":1435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":1362}},"tokens_in":575,"tokens_out":1435,"duration_ms":13399,"temperature":1.0,"reasoning_tokens":1362,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:16:47.289134+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Play a single pure tone near a container's resonant frequency as the only excitation, and evaluate the trained model on an unseen instance of that container: if the source-invariance claim is correct, the MAE should stay close to the 0.04 seen with ambient sound, while a large MAE increase would show that narrowband or notched excitations break the Fourier-magnitude proxy.","supporting_citations":[{"cited_title":"Visual vibrometry: Estimating material properties from small mo- tion in video","cited_arxiv_id":null,"evidence_quote":"Establishes that speaker-excited surface vibrations captured by camera can be used to estimate object properties; the methodological starting point for inferring liquid level from vibration."},{"cited_title":"Narasimhan","cited_arxiv_id":null,"evidence_quote":"The prior dual-shutter speckle vibration system that sensed a single row of points; this work extends the approach to a simultaneous 2D grid of points across multiple containers."},{"cited_title":"Simultaneous remote extraction of multiple speech sources and heart beats from secondary speckles pattern.Op- tics express, 17(24):21566–21580, 2009","cited_arxiv_id":null,"evidence_quote":"Demonstrates speckle-pattern shift as a measure of surface vibration from a defocused laser spot, the underlying sensing principle."},{"cited_title":"The phase correlation image alignment method","cited_arxiv_id":null,"evidence_quote":"Phase correlation method used to estimate integer-pixel speckle shifts between consecutive frames."},{"cited_title":"Attention is all you need","cited_arxiv_id":null,"evidence_quote":"Transformer architecture on which the Vibration Transformer's attention-based design is based."},{"cited_title":"Soft labels for ordinal regres- sion","cited_arxiv_id":null,"evidence_quote":"Introduces soft labels for ordinal regression, the basis of the SORD loss used for liquid-level classification."},{"cited_title":"Non-intrusive tank-filling sensor based on sound resonance","cited_arxiv_id":null,"evidence_quote":"Prior approach that infers liquid level from sound resonance after physically knocking on a container; the baseline the paper's remote visual method is designed to surpass."}],"review_version":1}