{"id":"221d70b3-9d38-49eb-9896-1ca8b0842681","arxiv_id":"2502.08869","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey organizing existing methods that encode time series as images and apply vision models, with a dual-view taxonomy of imaging and modeling approaches.","lead":"This survey catalogs methods that turn time series data into images and then use vision models to analyze those images. It offers a taxonomy of imaging techniques and model types, aiming to be a reference guide for researchers exploring vision-based time series analysis.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The survey's core category is underspecified: Table 1 includes methods that are not vision models (K-NN, SVM/AdaBoost, ConvLSTM), so the 'first survey on leveraging vision models' claim cannot be verified without a precise scope definition.","rationale":"Stress-test read: The reader's concern about missing search protocol is valid, but the deeper issue is that the paper never fixes the extension of its key term. A survey's 'first' and 'comprehensive' claims are only checkable if the surveyed class is well-defined. The paper defines the imaging side (line plots, heatmaps, spectrograms, GAF, RP) but not the modeling side. Table 1 includes at least three entries whose modeling components are not vision models: K-NN, SVM/AdaBoost ensembles, and ConvLSTM. This is not a mere labeling quibble: the inclusion of generic classifiers and a spatiotemporal sequence model makes the 'first survey on leveraging vision models' claim either trivially broad or inconsistently narrow. The paper's own exclusion of video/image-time-series methods (Section 2) shows the authors are aware of scope boundaries, yet no analogous boundary is set for what counts as a vision model. If the definition is 'any model that operates on images,' the survey should be titled 'image-based time series analysis' and would need to cover much earlier work; if the definition is 'models pre-trained on natural images,' then the conventional-vision section (Section 4.1) and many Table 1 rows should be removed or reframed. Neither is done, so the central claim cannot be assessed. That said, the survey has genuine value: the dual-view taxonomy of imaging methods is clearly organized, the pre/post-processing discussion (Section 5) is practical, and the Github resource is a concrete artifact. These positives are independent of the scope problem, so the verdict should remain CONDITIONAL rather than REJECT: the authors can fix the concern by adding a formal definition of 'vision model' and applying it consistently to Table 1.","tokens_in":16220,"tokens_out":8412,"duration_ms":82340,"concrete_test":"Apply a formal inclusion test to every row of Table 1: does the model in the 'Imaged Time Series Modeling' column belong to one of the classes defined in Section 4 (conventional vision model, LVM, LMM) under an explicit definition of 'vision model'? In particular, check whether K-NN ([Silva et al., 2013]), SVM/AdaBoost ([Cohen et al., 2020]), and ConvLSTM ([Zhang et al., 2019]) survive the definition. Re-run the taxonomy counts and the Section 1/7 'first survey' claim with the surviving set. If the surviving set substantially changes (rows removed or re-classified), the paper must either narrow its title to 'image-based time series analysis' or justify why generic classifiers count as vision models.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper bills itself as 'the first survey on leveraging vision models for time series analysis' and makes a comprehensive-coverage claim (Section 1, Section 7). The load-bearing precondition is that 'vision model' is a well-defined inclusion criterion on the modeling side. The paper never defines it. Section 2 defines only the time-series-to-image transformation side and then states an exclusion for 'methods on videos or sequential images.' The modeling taxonomy in Section 4.1 ('Conventional Vision Models') is where the ambiguity bites: Table 1 lists [Silva et al., 2013] with a K-NN classifier on recurrence plots, [Cohen et al., 2020] with an ensemble of SVM/AdaBoost on line plots, and [Zhang et al., 2019] with ConvLSTM on a sequence of correlation matrices. These are not vision models: K-NN, SVM, and AdaBoost are generic classifiers, and ConvLSTM is a spatiotemporal sequence model. If the intended scope is 'any model applied to imaged time series,' the category is so broad that the first-survey claim becomes trivial and almost certainly false, since image-based time series classification predates deep vision models (e.g., bitmaps, recurrence plots). If the intended scope is 'models developed for or pre-trained on natural images,' then those rows are out of scope and the coverage and comparative discussion in Section 4 must be revised. Neither interpretation is stated, so the central claim is unfalsifiable. This is more fundamental than the absent search protocol identified by the reader: even a systematic search would need a query concept that the paper does not supply.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper surveys methods that convert time series into image representations and then apply vision models for tasks such as classification, forecasting, and anomaly detection. It organizes the field with a dual-view taxonomy: imaging transformations (line plots, heatmaps, spectrograms, GAF, RP, and others) and modeling approaches (conventional vision models, LVMs, LMMs/VLMs, and task-specific heads). It also discusses pre-/post-processing considerations and future directions, and claims to be the first comprehensive survey on this topic.","tokens_in":16491,"tokens_out":7096,"duration_ms":69528,"significance":"If the scope is clarified, this will be a useful reference: the mathematical descriptions of the main imaging transforms are accurate, the comparative tables are informative, and the maintained GitHub repository is a practical asset. The survey does not present new experiments, so its value lies in organization and synthesis rather than empirical claims. The main risks are definitional: without a precise inclusion criterion for 'vision model' and a transparent literature collection protocol, the 'first comprehensive survey' claim is hard to assess.","major_comments":[{"comment":"The central claim that this is 'the first survey on leveraging vision models for time series analysis' requires a precise definition of 'vision model,' but none is given in the scope discussion. Section 2 defines only the imaging side and excludes video/image time series, without restricting the modeling side. As a result, Table 1 includes K-NN on recurrence plots [Silva et al., 2013], an SVM/AdaBoost ensemble on line plots [Cohen et al., 2020], and ConvLSTM on correlation-matrix images [Zhang et al., 2019], none of which is a vision model in the usual sense of a model developed for or pre-trained on images. If the intended scope is 'any classifier applied to imaged time series,' the first-survey claim becomes very broad and needs a supporting literature search; if it is 'models developed for natural images,' those rows and their comparative discussion in §4.1 must be removed or reframed. The inclusion criterion should be stated explicitly and the taxonomy and Table 1 adjusted accordingly.","section":"§2, Table 1"},{"comment":"The paper claims to be 'comprehensive' and 'the first survey' on this topic, but no systematic search protocol is described. Table 1 appears to be a curated list, and there is no statement of databases, search terms, inclusion/exclusion criteria, or time window used to assemble it. Without such a protocol, a reader cannot verify the completeness that the 'first comprehensive survey' claim depends on, and the risk of missing important methods (e.g., early bitmap-based approaches, which are mentioned only in passing in §3.6) remains. Please add a methodology paragraph describing how the literature was collected and how representative Table 1 is intended to be.","section":"§1, §7"}],"minor_comments":[{"comment":"The text refers to 'Fig. 1(b)' when discussing Heatmap for MTS, but the referenced heatmap is in Fig. 2(b); please correct the cross-reference.","section":"§3.7"},{"comment":"There is a typo in the phrase 'the deigns of task-specific heads'; 'deigns' should be 'designs.'","section":"§2.1"},{"comment":"In the Time Series Normalization paragraph, 'withtin [0, 1]' should be 'within [0, 1]'; the same typo appears in the GAF sentence.","section":"§5"},{"comment":"The term 'spaciotemporal' in the scope paragraph should be 'spatiotemporal.'","section":"§2"},{"comment":"The sentence 'ForCNN [Semenoglou et al., 2023] even uses a single 8-bit integer to represent each pixel for black-white images' reads awkwardly because 'ForCNN' is not formatted as a method name; please clarify whether it is a named method or a typo for 'For CNN.'","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the 'first survey' claim is plausible but cannot be verified without a search protocol, so I would ask the authors to add one and to tighten the scope of 'vision model.' Several works in Table 1 are authored by members of the survey team; this is not a problem in itself, but it makes transparency about the selection process particularly important."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first survey that actually organizes the 'turn a time series into an image and run a model on it' literature, and the dual taxonomy — imaging methods vs. modeling strategies — is a sensible way to do it. It is a survey, so no new method, but it ships something useful: a maintained GitHub repo with code for the imaging transforms, correct-looking math for STFT/CWT/GAF/RP, and a practical section on normalization, image alignment, and recovering time series from predicted images. That last part is the kind of thing people entering this area will genuinely appreciate.\n\nThe real soft spot is the scope definition. Section 2 says the survey covers methods that transform time series to images and then apply vision models. But Table 1 includes Silva et al. 2013 (K-NN on recurrence plots), Cohen et al. 2020 (SVM/AdaBoost on line plots), and MSCRED (ConvLSTM on correlation matrices). Those are not vision models. Either the survey means 'any model applied to imaged time series,' in which case the title and the 'first survey of vision models' claim need to be softened and the pre-deep-learning bitmap work cited, or those rows are out of scope and should be trimmed. The stress-test note calls this a load-bearing flaw making the first-survey claim unfalsifiable — I think that's overstated, because the paper does state a scope, it just doesn't stick to it. It's an inconsistency, not a missing concept. Still, it needs to be fixed before publication.\n\nThe other weakness, which the reader identified, is the unsubstantiated comprehensiveness claim. No search protocol is described, so 'comprehensive' is an assertion. A short paragraph on how the corpus was assembled would address it.\n\nCitation pattern is fine; some authors appear in a few cited papers, but the taxonomy doesn't lean on them. The math checks out. The per-method pros/cons table is a nice contribution. Who is this for: grad students and researchers deciding whether vision models are worth trying for their time series problem, and anyone wanting a catalog of imaging methods with their failure modes. It deserves a serious referee — the survey fills a real gap — but I'd ask for a tightened scope statement and a search methodology before accepting.","headline":"A genuinely useful first survey of imaging-based time series analysis, with a clear taxonomy and practical guidance; the loose definition of 'vision model' lets in K-NN and SVM rows that undercut the 'first survey of vision models' claim, but that is fixable with a sharper scope statement.","tokens_in":17045,"tokens_out":2715,"would_cite":true,"duration_ms":26811,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This survey claims to be the first comprehensive review of vision models for time series analysis, organizing the field with a dual-view taxonomy of imaging and modeling methods.","keywords":["time series analysis","vision models","large vision models","multimodal models","time series imaging","taxonomy","survey","forecasting"],"falsifier":"A systematic literature search with a defined protocol that finds a substantial number of vision-based time series methods not representable in the five imaging categories or the three modeling categories of Table 1 would weaken the survey's claim to comprehensiveness. Concretely, discovering a widely used imaging technique for multivariate series that is neither a Line Plot, Heatmap, Spectrogram, GAF, nor RP — or a modeling paradigm outside conventional CNNs, LVMs, and LMMs — would falsify the taxonomy's coverage.","tokens_in":16012,"feed_emoji":"📈","tokens_out":5364,"duration_ms":43012,"temperature":0.7,"pith_summary":"This paper is a survey that claims to be the first comprehensive review of using vision models for time series analysis. It argues that imaging time series and feeding the images to vision models is a promising alternative to sequence-based LLM approaches, because images naturally represent continuous values and can encode multivariate correlations. The paper organizes the field through two lenses: how time series are converted to images (line plots, heatmaps, spectrograms, Gramian Angular Fields, recurrence plots), and how the resulting images are modeled (conventional CNNs, large vision models, and multimodal models). It also collects practical guidance on normalization, image alignment, and recovering time series values from predicted images. A sympathetic reader would take away a structured map of the area and a set of open questions about when imaging helps.","feed_headline":"First survey maps how vision models read time series","feed_subtitle":"A dual taxonomy of imaging and modeling methods organizes the field and exposes open questions.","key_machinery":"The organizing mechanism is the 'imaged time series' pipeline depicted in Figure 1: time series preprocessing, transformation to an image, image preprocessing, modeling by a vision model, and post-processing recovery. The survey's taxonomy is the dual view of this pipeline: Time Series to Image Transformation (§3) and Imaged Time Series Modeling (§4). The five primary imaging methods are Line Plot (x-y plot of values over time), Heatmap (color-coded variate-time matrix), Spectrogram (time-frequency representation via STFT, wavelet, or filterbank), Gramian Angular Field (polar-coordinate Gram matrix encoding pairwise temporal correlations), and Recurrence Plot (pairwise distances of delay-embedded phase-space vectors). On the modeling side the survey distinguishes conventional vision models (CNNs, ResNets, GANs, diffusion U-Nets), large vision models (ViT, Swin, BEiT, MAE, often ImageNet-pretrained), and large multimodal models that accept both image and text prompts.","core_discovery":"The central claim is that the whole pipeline — from raw time series to image, through a vision model, and back to time series — is a coherent framework that deserves its own survey and taxonomy. The paper's distinctive contribution is the dual-view classification: one axis enumerates five primary imaging transforms (Line Plot, Heatmap, Spectrogram, Gramian Angular Field, Recurrence Plot) and their extensions to multivariate series; the other axis groups modeling strategies into conventional vision models, large vision models such as ViT and MAE, and large multimodal models such as LLaVA-based architectures. Within that scaffold, the survey collects findings such as VisionTS's one-epoch fine-tuned MAE becoming state-of-the-art forecaster and ViTime's image-pretrained ViT beating TimesFM in zero-shot forecasting. The paper also maps pre- and post-processing desiderata, including instance normalization, channel alignment for RGB inputs, and the varying ease of recovering a series from different image types.","pith_inferences":["The survey's framework suggests a natural benchmark study that fixes the imaging method and varies the vision model, or vice versa, across a common set of tasks; the survey itself notes that such systematic comparisons are largely absent.","The paper's emphasis on visual prompt-friendliness yields a testable extension: for long multivariate series, a plotted image plus a short text prompt should outperform a fully verbalized numeric sequence in zero-shot multimodal settings, at lower API cost.","One could push the 'images encode correlations' argument further by combining a heatmap of the variate-time matrix with a graph neural network over variate positions, which might address the row-order sensitivity the paper flags.","Because the survey claims to be the first of its kind, future surveys will need to update this taxonomy as multimodal and foundation models evolve; the two-axis structure may need a third axis for task types."],"forward_implications":["If the survey's framework is correct, researchers can slot any new vision-plus-time-series method into one of the two taxonomy axes and quickly identify what is genuinely new.","The reported results, such as MAE fine-tuned for one epoch becoming a state-of-the-art forecaster and ViTime beating TimesFM in zero-shot forecasting, suggest that image-based encoders are a credible rival to LLM-based time series models and deserve inclusion in future benchmarks.","The pre- and post-processing guidance implies that practitioners must check image size, channel count, and normalization before reusing pretrained vision models, and that some imaging formats (GAF) support exact recovery while others (RP) do not.","The survey's future-directions list points to vision-based time series foundation models and multimodal agents as the next targets for the field."],"supporting_citations":[{"why":"Provides the VisionTS result that visual masked autoencoders can serve as zero-shot time series forecasters, supporting the survey's claim that LVMs transfer to time series.","marker":"[Chen et al., 2024]"},{"why":"Presents ViTime, an image-pretrained ViT that reportedly beats TimesFM in zero-shot forecasting, a key evidence for the potential of vision models over LLM-based foundation models.","marker":"[Yang et al., 2024]"},{"why":"Establishes the canonical fine-tuning of an ImageNet-pretrained DeiT on audio spectrograms, a primary example in the LVM modeling category.","marker":"[Gong et al., 2021]"},{"why":"Introduces Gramian Angular Fields and the paradigm of encoding time series as images for CNN classification, foundational to the imaging taxonomy.","marker":"[Wang and Oates, 2015a]"},{"why":"Introduces Recurrence Plots, the basis of the RP imaging method and its subsequent uses in classification and forecasting.","marker":"[Eckmann et al., 1987]"},{"why":"Formulates the channel-independence assumption, which the survey contrasts with imaging methods that can explicitly encode multivariate correlations.","marker":"[Nie et al., 2023]"},{"why":"Questions the effectiveness of LLMs for time series forecasting, supporting the survey's motivation for seeking vision-model alternatives.","marker":"[Tan et al., 2024]"},{"why":"Demonstrates that plots unlock time-series understanding in multimodal models, supporting the advantage of visual prompting over verbalized numeric sequences.","marker":"[Daswani et al., 2024]"},{"why":"Shows image transforms with diffusion models for generative time series modeling, supporting the 'Other methods' category and the compact encoding of long series.","marker":"[Naiman et al., 2024]"}],"fun_headline_variants":["Vision models beat LLMs for time series: a survey","Vision models read time series: dual taxonomy survey","Dual taxonomy for vision-based time series analysis","Imaging time series: a survey of vision models","Survey: vision models gain edge over LLMs for time series"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's claim to be comprehensive and its taxonomy's utility rest on the assumption that the papers collected in Table 1 form a representative and sufficiently complete sample of vision-based time series methods; the paper does not describe a systematic search protocol for the literature.","fun_headline_variants_meta":{"raw":{"variants":["Vision models beat LLMs for time series: a survey","Vision models read time series: dual taxonomy survey","Dual taxonomy for vision-based time series analysis","Imaging time series: a survey of vision models","Survey: vision models gain edge over LLMs for time series"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001311,"raw_usage":{"total_tokens":5333,"prompt_tokens":926,"completion_tokens":4407,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":4330}},"tokens_in":542,"tokens_out":4407,"duration_ms":29235,"temperature":1.0,"reasoning_tokens":4330,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T23:22:13.937141+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A systematic literature search with a defined protocol that finds a substantial number of vision-based time series methods not representable in the five imaging categories or the three modeling categories of Table 1 would weaken the survey's claim to comprehensiveness. Concretely, discovering a widely used imaging technique for multivariate series that is neither a Line Plot, Heatmap, Spectrogram, GAF, nor RP — or a modeling paradigm outside conventional CNNs, LVMs, and LMMs — would falsify the taxonomy's coverage.","supporting_citations":[{"cited_title":"Ast: Audio spectrogram transformer","cited_arxiv_id":null,"evidence_quote":"Establishes the canonical fine-tuning of an ImageNet-pretrained DeiT on audio spectrograms, a primary example in the LVM modeling category."},{"cited_title":"Recurrence plots of dynamical systems","cited_arxiv_id":null,"evidence_quote":"Introduces Recurrence Plots, the basis of the RP imaging method and its subsequent uses in classification and forecasting."},{"cited_title":"A time series is worth 64 words: Long- term forecasting with transformers","cited_arxiv_id":null,"evidence_quote":"Formulates the channel-independence assumption, which the survey contrasts with imaging methods that can explicitly encode multivariate correlations."},{"cited_title":"Are language models ac- tually useful for time series forecasting? In NeurIPS,","cited_arxiv_id":null,"evidence_quote":"Questions the effectiveness of LLMs for time series forecasting, supporting the survey's motivation for seeking vision-model alternatives."},{"cited_title":"Utilizing image transforms and diffusion models for generative modeling of short and long time series","cited_arxiv_id":null,"evidence_quote":"Shows image transforms with diffusion models for generative time series modeling, supporting the 'Other methods' category and the compact encoding of long series."}],"review_version":1}