{"id":"01920e48-2322-4b34-9ead-4ce15b94d894","arxiv_id":"2501.13066","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The SMART-Vision survey organizes vision-based human action recognition into a hybrid Venn-diagram taxonomy and reviews the emerging open-set/open-world HAR literature.","lead":"This paper is a survey that organizes deep learning methods for human action recognition into a new taxonomy, SMART-Vision, which highlights how methods combine into hybrids. It also reviews the emerging area of Open-HAR, where systems must handle action classes never seen during training.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark tables 9–11 contain verified dataset/split misassignments, so the survey's performance-comparison contribution is not currently trustworthy.","rationale":"The reader's weakest assumption is exactly where I would place the load-bearing risk. The paper's novel contribution is organizational, not empirical, so its value is entirely in accurate curation. The identified errors are not cosmetic: X3D's 94.6% under HMDB-51 would put it far above the actual state of the art, misleading any reader using the table to choose a model; applying NTU split terminology to Kinetics results indicates the numbers were copied across rows without verification. A table audit is cheap, decisive, and directly tests the survey's reliability. The Open-HAR 'first' claim is also unsupported by citation of prior surveys, but it is secondary and can be fixed by rewording. Therefore the verdict should remain CONDITIONAL until the audit is done.","tokens_in":44102,"tokens_out":4730,"duration_ms":48067,"concrete_test":"Write a small audit script: for every model row in Tables 9–11, locate the cited paper's abstract or results table, record the reported accuracy and dataset/split, and compare. Flag every mismatch, including missing percent signs and wrong columns. In particular, verify whether X3D's 94.6% appears in the X3D paper as Kinetics-400 (not HMDB-51), and whether TubeViT-H/UMT-L are ever reported with NTU XSub/XView splits (they are Kinetics models). If the mismatch rate is nonzero, recompute any comparative claims (e.g., 'best on Kinetics,' 'state-of-the-art on NTU') and list corrections before the survey is cited as a reference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The survey's stated value is a reliable map: a taxonomy plus a 'comprehensive evaluation and performance comparison' (Section 3.9). That map is only as good as the transcribed numbers and split labels. The reader found the first cracks, and they are not isolated typos. Table 9 lists X3D as 94.6% under HMDB-51, while Section 3.9 explicitly attributes 94.6% to Kinetics-400. Section 3.9 assigns NTU RGB+D 'XSub/XView' split names to Kinetics-400/600 results for TubeViT-H and UMT-L, and Table 11 repeats the mislabeling. Table 13 lists HMDB51 twice with contradictory modality flags (once RGB-only, once RGB+S+D+IR), and Table 14 duplicates UCFKinect. These are data-integrity failures in the comparative core of the paper. If the benchmark tables cannot be trusted, the roadmap and the hybrid taxonomy lose their evidentiary basis; the central claim that many systems are 'inherently hybrid' rests on paper assignments that are not independently checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a survey of deep-learning-based vision approaches to human action recognition (HAR). Its central contribution is a 'SMART-Vision' taxonomy, displayed as a Venn diagram, that groups HAR architectures into two-stream, 3D convolutional, graph convolutional, motion, and Transformer families, and emphasizes that many recent systems are hybrids of these families. The survey also claims to provide the first comprehensive analysis of Open-HAR (open-set and open-world HAR), a comparison of benchmark results in Tables 9-11, a review of HAR datasets, and a discussion of challenges and future directions.","tokens_in":44296,"tokens_out":3525,"duration_ms":36970,"significance":"If the taxonomy and literature mapping are accurate, the survey is a useful organizing resource for a fast-moving field: the Venn-diagram structure and the per-intersection citation tables give a practical map of how 2D/3D CNNs, GCNs, motion modules, and Transformers are combined. The Open-HAR section is a valuable collection of scattered work on open-set and open-world HAR, including the KOWL-718 protocol, and it is more detailed than in most prior HAR surveys. The paper does not claim a new empirical result or benchmark, so its value rests on the reliability of its literature assignments and tables. That reliability is currently compromised by verified transcription and labeling errors in the comparative core (Tables 9, 11, 13, 14), which must be corrected before the survey's 'comprehensive evaluation and performance comparison' claim can be accepted.","major_comments":[{"comment":"Table 9 lists X3D as achieving 94.6% under the HMDB-51 column, while the text in Section 3.9 explicitly states that this value is a Kinetics-400 accuracy. The same row leaves the K400 column blank. Because Table 9 is the evidence base for the paper's 'comprehensive evaluation and performance comparison' contribution, this is a load-bearing data-integrity error, not a formatting slip. All numbers in Tables 9-11 should be re-verified against the primary sources and placed in the correct dataset columns.","section":"Table 9 / Section 3.9"},{"comment":"The paragraph discussing TubeViT-H and UMT-L reports Kinetics-400 and Kinetics-600 Top-1/Top-5 accuracies using NTU-style split names: 'TubeViT-H model produced 90.9% accuracy for the XSub and 98.9% the XView on Kinetics-400' and 'UMT-L Model provide 98.8% for XView on Kinetics-600.' Table 11 repeats these values as Top-1/Top-5, so the text contradicts its own table and mislabels standard Kinetics metrics as cross-subject/cross-view splits. The section should clearly separate Kinetics Top-1/Top-5 results from NTU RGB+D XSub/XView results.","section":"Section 3.9 / Table 11"},{"comment":"Table 13 lists HMDB51 twice with contradictory modality entries: the first entry marks it as RGB-only, while a later entry for the same dataset marks RGB, Skeleton, Depth, and Infrared as available. This second row contradicts the dataset description in Section 5, where HMDB51 is presented as an RGB dataset. Since the table is the survey's modality-coverage reference, the duplicated row and its checkmarks must be reconciled with the primary dataset documentation.","section":"Table 13"},{"comment":"Table 14 contains duplicate rows for UCFKinect, with sample counts written as '12,80' (presumably 1,280), and it labels Kinetics-400/600/700 as 'Kinectis' throughout. These entries, together with the errors in Tables 9 and 13, indicate that the tabulated inventories were not carefully checked against the cited sources; a systematic verification pass is needed before the survey can serve as a reliable reference.","section":"Table 14"}],"minor_comments":[{"comment":"The text contains 'ST-GNC' where ST-GCN is meant, and 'different mythologies' where 'methodologies' is intended; these typos should be corrected.","section":"Section 3.4.3"},{"comment":"The RHN row reports '71.8' without a percent sign, unlike all other entries; this is likely a typographical omission, but every numerical cell should be checked for consistency.","section":"Table 9"},{"comment":"The caption abbreviates '3D Convolutional Networks' as '3DCN', while the text immediately below refers to 'Graph Convolutional Networks (3DCN)', creating an ambiguous acronym; the abbreviations for the two network families should be made distinct.","section":"Figure 1 caption"},{"comment":"The sentence 'Other researchers have contributed exciting to the two-stream architecture' appears to be missing a noun or adjective; it should read 'exciting contributions' or similar.","section":"Section 3.2.2"},{"comment":"The Kinetics series is consistently misspelled as 'Kinectis' in Table 14 and in parts of Section 5; the correct spelling should be used throughout.","section":"Section 5, Table 14"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's header states 'Published in Multimedia Tools and Applications Journal' while the arXiv submission is dated January 2025; the editor may wish to verify the publication status and ensure that posting to arXiv is consistent with the journal's policies. The self-citations in the Open-HAR and applications sections are topical and do not appear to inflate the paper's central claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this survey has a genuinely useful organizing idea, but its performance tables are not reliable as printed. The SMART-Vision taxonomy—five architectural families plus hybrid intersections, with tables assigning papers to each intersection—is a real contribution. The Open-HAR section is a decent entry point to a scattered literature. But the benchmark comparison, which the paper advertises as a core contribution, is currently compromised by verified transcription and labeling errors.\n\nThe stress-test holds up on reading the paper. Table 9 places X3D at 94.6% under HMDB-51, while Section 3.9 explicitly says that number is a Kinetics-400 result. Section 3.9 also assigns NTU RGB+D XSub/XView split names to Kinetics-400 results for TubeViT-H and UMT-L, and Table 11 repeats the mislabeling. Table 13 lists HMDB51 twice with contradictory modality columns, and Table 14 duplicates UCFKinect. These are not isolated typos; they are data-integrity failures in the comparative core. The taxonomy's paper-to-category assignments may also be affected by the same carelessness, though I did not verify each one.\n\nWhat the paper does well: the hybrid-intersection view is a meaningful improvement over earlier taxonomies that treat architectures as monolithic. The discussions per family are informative, and the Open-HAR review usefully collects known/unknown split protocols, metrics, and the KOWL-718 baseline, and flags the metric problems itself. That part is worth reading.\n\nSoft spots in proportion: besides the tables, the 'first comprehensive analysis of Open-HAR' claim is asserted without benchmarking against prior surveys—tone that down. Some self-citation in the Open-HAR section (Boult, Prijatelj, and the authors' own works) is present, but it does not drive the taxonomy, so it is minor. The survey is long and would benefit from tighter editing, but that is minor.\n\nBottom line: the taxonomy and Open-HAR overview deserve referee time, but only after a careful pass over every table entry. I would not cite the benchmark numbers in their current state. The paper is best for newcomers wanting a roadmap, not for anyone seeking reliable accuracy comparisons.\n\nRecommendation: accept for peer review with major revision; require a full re-verification of all tables against original papers, correction of split labels, and removal or qualification of the 'first' claim. I'd read a revised version.","headline":"A useful hybrid-taxonomy and Open-HAR overview, but the benchmark tables have verified errors that currently undermine the paper's comparative claims.","tokens_in":44860,"tokens_out":2086,"would_cite":false,"duration_ms":22634,"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":"A new survey taxonomy, SMART-Vision, argues that most modern human-action-recognition systems are hybrids of two-stream, 3D-convolutional, graph-convolutional, motion, and transformer designs, and gives a first overview of the emerging…","keywords":["human action recognition","deep learning","hybrid architectures","two-stream networks","3D convolutional networks","graph convolutional networks","vision transformers","open-world recognition"],"falsifier":"Check each entry of Tables 9-11 against the cited original papers: if entries such as X3D's 94.6%, TubeViT-H's XSub and XView figures, and UMT-L's XView figure are confirmed to be assigned to the wrong datasets or values, the comprehensive evaluation and performance comparison contribution is not reliable; a full re-verification would settle the question.","tokens_in":43908,"feed_emoji":"🎥","tokens_out":6165,"duration_ms":53339,"temperature":0.7,"pith_summary":"This survey paper sets out to show that the field of vision-based human action recognition (HAR) has moved beyond clean architectural camps: most modern systems combine two-stream, 3D-convolutional, graph-convolutional, motion, and transformer ideas. Its instrument is the SMART-Vision taxonomy, a Venn-style map whose intersection tables assign individual papers to the hybrid combinations they actually use. The paper also claims to give the first comprehensive survey of Open-HAR, an emerging evaluation setting in which systems must handle action classes never seen at training time. A sympathetic reader would care because the taxonomy promises a usable route through a literature that has become entangled across paradigm shifts, and it identifies open-set and open-world evaluation as a gap the field has not yet standardized. The survey's value is organizational: it is a map and a baseline inventory rather than a new empirical result.","feed_headline":"One taxonomy maps how action recognition went hybrid","feed_subtitle":"SMART-Vision traces five deep-learning shifts and the first survey of open-world action recognition.","key_machinery":"The central object is the SMART-Vision taxonomy: a Venn diagram (Figure 1) with five architectural families as overlapping regions and a Hybrid Networks region formed by their unions, accompanied by per-intersection tables (Tables 1-8, 9-11) that list representative papers for each combination. It is the device that carries the argument, because each citation's placement in an intersection is the evidence that the system is hybrid. The second machinery is the Open-HAR framing, which imports the open-set and open-world distinction from general recognition into HAR: open-set systems reject unknown classes, while open-world systems detect, cluster, and incrementally learn them. The taxonomy's claim that many proposed systems are inherently hybrid is operationalized entirely through these placements.","core_discovery":"On the paper's own terms, the discovery is that hybridity is the norm rather than the exception in modern HAR. The authors organize the literature into five foundational families — two-stream networks, 3D convolutional networks, graph convolutional networks, motion networks, and transformer networks — and then show, paper by paper, that the most successful systems occupy the intersections: two-stream I3D couples motion with 3D convolution, ST-TR and RGBSformer pair skeletons with transformer attention, and SlowFast-style designs merge dual-rate streams with 3D backbones. For Open-HAR, the paper assembles the scattered open-set and open-world work, from evidential and capsule-based open-set classifiers to the KOWL-718 open-world protocol, into a single discussion with the explicit claim that no accepted large-scale evaluation standard yet exists. The claim is therefore about how to read the field: a fair reader should see modern action recognition as a web of complementary design choices, not a race between monolithic architectures.","pith_inferences":["If the hybrid map is right, performance comparisons should routinely ablate each component of a hybrid system; the survey does not itself provide such ablations, so readers cannot tell which ingredient carries the gain.","The same Venn-style mapping could be applied to neighboring video tasks, such as temporal action localization, video-text retrieval, and egocentric action recognition, where similar paradigm blending is underway but not yet cataloged.","Open-HAR's lack of a standard evaluation protocol suggests a concrete next step: re-run existing open-set HAR methods on the KOWL-718 splits using OWM-style metrics, which the survey notes have not yet been adopted.","Because the comparison tables already contain at least one dataset-column mismatch, the benchmark inventory should be treated as a pointer to original papers rather than as a verified leaderboard."],"forward_implications":["A newcomer can use the intersection tables to find the hybrid recipe that matches a task, such as skeleton plus transformer or two-stream plus 3D convolution, instead of reading the full literature.","If hybridity is the norm, the field's next advances are likely to come from combining existing families, such as GCN pose encoders with masked-autoencoder transformers, rather than from a new monolithic architecture.","Open-HAR becomes a testable research program: with the KOWL-718 protocol available, future open-world HAR systems can be compared on a common large-scale baseline.","The dataset and benchmark tables give a single place to compare representative accuracy figures across UCF-101, HMDB-51, Kinetics, and NTU RGB+D, subject to verification of the transcribed numbers.","The survey's challenges list points to occlusion, viewpoint, efficiency, and generalization as the binding constraints on deployed HAR systems, not classification accuracy alone."],"supporting_citations":[{"why":"Defines the two-stream architecture that anchors the taxonomy's T-SN category.","marker":"[25]"},{"why":"Provides two-stream I3D, the canonical hybrid of two-stream and 3D-convolutional designs.","marker":"[55]"},{"why":"Introduces SlowFast, a dual-rate two-stream and 3D design used throughout the survey's hybrid discussion.","marker":"[46]"},{"why":"Introduces ST-GCN, the foundational skeleton graph convolutional network for the GCN category.","marker":"[70]"},{"why":"Originates the Transformer and self-attention machinery underlying the transformer category.","marker":"[120]"},{"why":"Introduces the Vision Transformer used by many HAR models the survey classifies.","marker":"[49]"},{"why":"Presents ST-TR, a skeleton spatial-temporal transformer at the GCN and transformer intersection.","marker":"[127]"},{"why":"Defines open-world recognition and the three required system components for the Open-HAR section.","marker":"[194]"},{"why":"Formalizes Open-world HAR and provides the KOWL-718 large-scale protocol.","marker":"[198]"},{"why":"Presents DEAR, the first evidential open-set HAR system around which several Open-HAR evaluations are built.","marker":"[209]"}],"fun_headline_variants":["Hybrid models rule action recognition, survey shows","Action recognition's secret: hybrid designs win","From two-stream to transformers: hybrid HAR survey","Hybridity is the new normal in action recognition","Survey puts hybrid action recognition in the spotlight"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise of the survey's comparative claims is that the accuracy numbers transcribed into Tables 9-11 are correct and assigned to the right dataset columns; at least one entry violates this premise, since Table 9 lists X3D's 94.6% under HMDB-51 while the text describes it as a Kinetics-400 result.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid models rule action recognition, survey shows","Action recognition's secret: hybrid designs win","From two-stream to transformers: hybrid HAR survey","Hybridity is the new normal in action recognition","Survey puts hybrid action recognition in the spotlight"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000771,"raw_usage":{"total_tokens":3417,"prompt_tokens":953,"completion_tokens":2464,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":2394}},"tokens_in":569,"tokens_out":2464,"duration_ms":18668,"temperature":1.0,"reasoning_tokens":2394,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:28:35.530461+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check each entry of Tables 9-11 against the cited original papers: if entries such as X3D's 94.6%, TubeViT-H's XSub and XView figures, and UMT-L's XView figure are confirmed to be assigned to the wrong datasets or values, the comprehensive evaluation and performance comparison contribution is not reliable; a full re-verification would settle the question.","supporting_citations":[],"review_version":1}