{"id":"ceaedb36-b0a7-4c05-b4a8-62e96f7d3ea1","arxiv_id":"2507.22061","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MOVE provides a new motion-guided few-shot video object segmentation benchmark, and the proposed DMA baseline outperforms six existing methods across all settings.","lead":"This paper introduces MOVE, a dataset of 4,300 videos with motion-defined categories for few-shot video object segmentation, plus a baseline model called DMA. The benchmark shows existing category-based methods struggle on motion-guided tasks, and DMA achieves the best reported results.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark's motion/category decorrelation is unverified: without per-motion object-category statistics, the DMA-vs-baseline gaps in Tables 3-4 could reflect category shortcuts rather than motion understanding.","rationale":"Read in good faith: the paper's contribution is a new benchmark plus a DMA baseline, and the benchmark's validity requires that motion, not object category, is the reliable cue in the evaluation episodes. The paper gives qualitative examples and aggregate statistics, but not the joint distribution of object and motion categories. This is the same weakest assumption the reader identified, and I agree it is load-bearing. I do not see a separate problem of equal weight: the DMA results do consistently dominate Tables 3-4, the ablations are internally consistent, and the proposed matching score is a sensible mechanism. The unresolved point is whether the observed gaps prove motion understanding or reflect category co-occurrence. A category-prior control and per-motion object-category entropy would directly settle this, and the paper should provide them before the motion-isolation claim is accepted. The verdict remains CONDITIONAL rather than REJECT because the dataset and baseline retain independent value if the confound is controlled, but the central interpretation is not yet fully supported.","tokens_in":15052,"tokens_out":5341,"duration_ms":69291,"concrete_test":"Compute the conditional matrix P(object category | motion category) over MOVE's 5,135 objects and the episode-level distribution for the OS and NS evaluation episodes, reporting the average entropy H(object | motion) weighted by motion frequency and the fraction of 2-way episodes in which the query target's object category appears in exactly one support video. Then train a category-only control baseline: use the support videos' object-category labels with a pretrained open-vocabulary segmenter (e.g., SAM2 with CLIP prompts) to segment the query object whose category matches the support, with no motion features, and report J&F, T-Acc, N-Acc on the same episodes.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim, that MOVE isolates motion as the discriminative signal, rests on the assumption that motion categories are not confounded with object categories. Section 3.3 reports only aggregate counts (224 motions, 88 object categories, 5,135 objects); it never reports P(object | motion) or the episode-level co-occurrence of object categories between support and query. If most motions are dominated by one object type (e.g., sports ≈ person), a model can segment the query by category prior without using temporal cues. In the 2-way setting, a still stronger shortcut exists: whenever the two support videos have different object categories and the query target matches the object category of one support, category matching picks the right candidate even though it ignores motion. Table 6 already shows appearance-only prototypes reach 36.5 J&F (vs. 46.8 combined), so appearance/category cues carry substantial signal; the paper does not decompose how much of that is legitimate instance identity versus category leakage. The necessity study (Table 2) and the main comparisons therefore do not yet establish that the observed performance gaps on MOVE are caused by motion understanding; they could be caused by the dataset's object-category structure, especially since the NS split is described only at the parent-class level in the supplementary.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MOVE, a large-scale benchmark for motion-guided few-shot video object segmentation (FSVOS), where support and query videos are matched by motion pattern rather than object category, and introduces DMA (Decoupled Motion-Appearance Network), a baseline that extracts separate appearance and motion prototypes with auxiliary classification supervision, prototype attention, and a mask decoder. The authors evaluate six prior methods from FSVOS, few-shot image segmentation, and referring video object segmentation on MOVE under overlapping and non-overlapping splits, in 2-way-1-shot and 5-way-1-shot settings with ResNet50 and VideoSwin-T backbones, and report that DMA outperforms all baselines on J&F and T-Acc, though all methods achieve very low N-Acc. The paper also reports a necessity study, ablations of the motion extractor and prototype decoupling, oracle experiments, and qualitative examples.","tokens_in":15290,"tokens_out":3015,"duration_ms":39505,"significance":"If the benchmark's motion-isolation claim holds, MOVE would be a valuable new resource: it is larger and more motion-focused than existing FSVOS datasets, uses video-level support sets, and provides a concrete task definition that could catalyze research on motion-centric few-shot segmentation. The DMA method is a reasonable first baseline, and the oracle experiments give useful upper bounds. The internal ablations are consistent and the benchmarking effort is broad. However, the central claim that MOVE isolates motion as the discriminative signal is currently under-supported because the dataset statistics do not rule out strong correlations between motion categories and object categories, and the baseline adaptations are not described in sufficient detail to allow reproduction or to rule out benchmarking bias. The low N-Acc values also temper the claim of consistent superiority across all metrics.","major_comments":[{"comment":"","section":"§3.3 and Table 2"},{"comment":"","section":"§5.2, Tables 3-4"},{"comment":"","section":"§5, Implementation Details and baseline adaptation"},{"comment":"","section":"§5.1, Table 2, HPAN*"}],"minor_comments":[{"comment":"","section":"§5.2, paragraph 2"},{"comment":"","section":"§4.3, Eq. (3)"},{"comment":"","section":"§5.3, Table 7"},{"comment":"","section":"§3.3 and NS split description"},{"comment":"","section":"§5.4, Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The dataset contribution is potentially strong and the DMA method is a sensible baseline, but the paper's central claim—that MOVE isolates motion from object category—is not yet supported by the presented statistics. The addition of per-motion object-category distributions and a controlled experiment removing category leakage would be decisive. I would also urge the editor to require the authors to release the adapted baseline code and evaluation scripts, since the paper's benchmarking conclusions depend on them. No concern about novelty or scope; the issue is verification and reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Jia,\n\nQuick take: this is a benchmark paper worth taking seriously, and the dataset is the real contribution. MOVE is new—224 motion categories, video-level support sets, 4,300 videos, over 300k masks. That is a substantial annotation effort. The DMA baseline is incremental but clean: decoupled motion and appearance prototypes with auxiliary classification, and the ablations show the motion branch helps. The benchmarking of six methods from three task families is also useful.\n\nThe soft spot is exactly the one flagged in the stress test: the paper never shows that motion categories are decorrelated from object categories. It reports 224 motions and 88 object categories but no P(object | motion) or episode-level co-occurrence. If most sports motions involve people, a model can solve the task by segmenting the dominant semantic class. The 2-way setting makes this worse—if the two support videos contain different object categories, category matching alone picks the correct target. Table 6 shows appearance-only prototypes reach 36.5 J&F, so appearance carries real signal; the paper does not decompose how much of that is legitimate instance identity versus category leakage. This is a load-bearing issue for the benchmark's stated purpose, not a footnote.\n\nOther issues are proportionally smaller but real: no error bars or significance tests; N-Acc is very low for every method including DMA (11.5%), which the authors acknowledge but do not address; the baseline adaptations are only in the supplementary; code and data are not released; and there is a typo in Table 8 (\"Orcale\"). The oracle experiments are a good addition, though they also show the model has large headroom.\n\nNone of this kills the paper. The MOVE task is worth having, and the dataset can be made valid with added statistics: report per-motion object distributions, show example motions performed by multiple object types, and rerun or control the evaluation for category-balanced episodes. The DMA method is fine as a baseline.\n\nWho benefits: researchers in few-shot video segmentation, video retrieval, and motion understanding. A serious referee should engage, but the revision needs to address the confound and release the data before the central \"motion understanding\" claim is convincing.\n\nRecommendation: send to peer review, with a request for major revision.","headline":"The MOVE dataset is a genuine new resource for motion-guided few-shot video segmentation, but the paper's central claim that it isolates motion from object category is not yet backed by dataset statistics.","tokens_in":15815,"tokens_out":1879,"would_cite":true,"duration_ms":26679,"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 introduces MOVE, a dataset and task in which a few support videos of a motion specify what to segment, and reports that its decoupled motion-appearance network outperforms all six compared methods on it.","keywords":["few-shot video object segmentation","motion-guided segmentation","video object segmentation","motion understanding","prototype learning","decoupled motion-appearance network","MOVE benchmark"],"falsifier":"Compute the per-motion distribution of object categories in MOVE; if most motion classes involve a single object category, and a baseline that simply segments that category reaches a J&F comparable to DMA's, the benchmark would be measuring category matching rather than motion understanding.","tokens_in":14850,"feed_emoji":"🎬","tokens_out":9181,"duration_ms":94884,"temperature":0.7,"pith_summary":"This paper introduces MOVE, a few-shot video object segmentation benchmark in which the support set is a video-mask pair demonstrating a motion, and the goal is to segment every object in query videos performing that same motion regardless of object category. The dataset contains 224 motion categories, 4,300 videos, and more than 314,000 annotated masks. The paper reports that category-centric methods from image segmentation, video segmentation, and referring segmentation all lose accuracy on MOVE compared with category-based benchmarks, which it reads as evidence that current models do not truly exploit temporal motion. The proposed DMA baseline extracts appearance and motion prototypes separately and is reported to outperform all six evaluated baselines across all settings and backbones, with 50.1% versus 45.4% J&F in the 2-way-1-shot overlapping-split setting. If the claim holds, MOVE provides the field with a new task, a reference baseline, and a demonstration that motion can be the discriminative signal in few-shot segmentation.","feed_headline":"New benchmark segments video objects by motion, not category","feed_subtitle":"Dataset of 224 motion classes plus a decoupled motion-appearance baseline beats all tested few-shot methods.","key_machinery":"The load-bearing object is the Decoupled Motion-Appearance module (DMA), which converts a support video and its mask into two separate prototype sequences. The appearance prototype is produced by mask-pooling per-frame features, so it encodes what the object looks like, while the motion prototype is produced by temporally differencing adjacent-frame features and enhancing them with 3D convolutions, so it encodes how the object moves. Two auxiliary classification heads, one for object categories and one for motion categories, keep the prototypes separated, and a transformer with cross-attention refines them before the mask decoder. This decoupling is the mechanism that lets a query object be matched to a support motion even when the two objects belong to different categories.","core_discovery":"The central claim is that temporal motion, not object category, can serve as the supervised signal for few-shot video object segmentation, and that a benchmark built on this idea exposes a real weakness in existing methods. MOVE changes the support set from static images to video clips with mask sequences, so the model must infer a motion prototype across frames, and the query set includes objects performing the support motion among distractors and empty frames. The paper's DMA network computes an appearance prototype by mask-pooling frame features and a motion prototype by differencing adjacent-frame features, keeps the two prototypes decoupled through separate auxiliary classification heads, and fuses them with transformer attention before decoding masks. On MOVE, DMA is reported to outperform all baselines in both 2-way-1-shot and 5-way-1-shot settings, on overlapping and non-overlapping motion splits, and with ResNet50 and VideoSwin-T backbones. The paper also shows that adding frame-level temporal modeling to a strong category-based baseline raises its J&F from 44.4% to 46.3%, supporting the claim that motion is the active ingredient.","pith_inferences":["A check the paper leaves implicit is to report the conditional distribution of object categories within each motion class; if most motions are performed by one object type, a model could solve the task by category matching alone.","The same episode construction could extend to other dense prediction settings, such as few-shot video instance segmentation or motion-guided video object detection, where the support videos would define a motion rather than a class.","The paper lists decomposing motions into primitive units as future work; if supported, that direction would test whether motion prototypes learned on known motions transfer to unseen motion combinations.","The matching-score head could be evaluated as a stand-alone few-shot action retrieval signal, since the paper motivates motion-based retrieval but reports no retrieval experiment."],"forward_implications":["Existing few-shot video segmentation methods will need explicit temporal modeling to perform well on MOVE; the paper reports that adding a simple self-attention step across frames raises a category-based baseline from 44.4% to 46.3% J&F.","Motion prototypes alone outperform appearance prototypes alone on MOVE (43.8% versus 36.5% J&F), so dynamic cues appear to carry most of the discriminative signal in this task.","Because DMA beats all six baselines under both backbones and both data splits, it can serve as a stable reference baseline for future work on MOVE.","The non-overlapping split is harder than the overlapping split for every method, meaning generalization to motion families that share no parent class with training is a distinct open challenge.","Low N-Acc values across all methods show that rejecting empty query frames and suppressing false positives is a common weakness, pointing to background modeling as a needed direction."],"supporting_citations":[{"why":"Defines the original few-shot video object segmentation task and provides the DANet baseline that is the strongest category-centric comparison for DMA.","marker":"[3]"},{"why":"Supplies the HPAN baseline that the paper adapts into HPAN* with temporal self-attention to show temporal modeling helps on MOVE.","marker":"[54]"},{"why":"Supplies the referring-video-segmentation baseline LMPM from the MeViS work; its performance drops sharply when text expressions are replaced by motion categories.","marker":"[11]"},{"why":"Supplies the TTI few-shot video object segmentation baseline that is among the strongest competitors on the non-overlapping split.","marker":"[51]"},{"why":"Supplies the SCCAN image-based few-shot segmentation baseline used in the necessity study, where its performance falls from 62.3% on YouTube-VIS to 40.6% on MOVE.","marker":"[59]"},{"why":"Supplies the CyCTR few-shot segmentation baseline that transfers poorly to MOVE, representing the image-centric approach in the benchmark tables.","marker":"[69]"},{"why":"Provides YouTube-VIS as the category-based comparison dataset in the necessity study.","marker":"[63]"},{"why":"Provides Kinetics-400 pretraining for the VideoSwin-T backbone, whose temporal features improve DMA's results in the main tables.","marker":"[24]"}],"fun_headline_variants":["Motion, not category, drives video object segmentation","New benchmark and baseline for motion-guided video segmentation","MOVE dataset: segment by motion, not object type","Decoupling motion and appearance wins few-shot video segmentation","Motion-guided few-shot video segmentation: benchmark and DMA baseline"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark assumes that in MOVE the same motion is performed by many different kinds of objects, so a model cannot solve the task by recognizing the dominant object category, such as always segmenting the human, instead of understanding the motion.","fun_headline_variants_meta":{"raw":{"variants":["Motion, not category, drives video object segmentation","New benchmark and baseline for motion-guided video segmentation","MOVE dataset: segment by motion, not object type","Decoupling motion and appearance wins few-shot video segmentation","Motion-guided few-shot video segmentation: benchmark and DMA baseline"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1374,"prompt_tokens":918,"completion_tokens":456,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":380}},"tokens_in":534,"tokens_out":456,"duration_ms":5098,"temperature":1.0,"reasoning_tokens":380,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:03:36.758335+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the per-motion distribution of object categories in MOVE; if most motion classes involve a single object category, and a baseline that simply segments that category reaches a J&F comparable to DMA's, the benchmark would be measuring category matching rather than motion understanding.","supporting_citations":[{"cited_title":"Delving Deep Into Many-to-Many Attention for Few-Shot Video Object Segmentation","cited_arxiv_id":null,"evidence_quote":"Defines the original few-shot video object segmentation task and provides the DANet baseline that is the strongest category-centric comparison for DMA."},{"cited_title":"Holistic prototype attention network for few-shot video object segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the HPAN baseline that the paper adapts into HPAN* with temporal self-attention to show temporal modeling helps on MOVE."},{"cited_title":"MeViS: A large-scale benchmark for video segmentation with motion expressions","cited_arxiv_id":null,"evidence_quote":"Supplies the referring-video-segmentation baseline LMPM from the MeViS work; its performance drops sharply when text expressions are replaced by motion categories."},{"cited_title":"Temporal transductive inference for few- shot video object segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the TTI few-shot video object segmentation baseline that is among the strongest competitors on the non-overlapping split."},{"cited_title":"Self-calibrated cross attention network for few-shot segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the SCCAN image-based few-shot segmentation baseline used in the necessity study, where its performance falls from 62.3% on YouTube-VIS to 40.6% on MOVE."},{"cited_title":"Few-shot segmentation via cycle-consistent trans- former","cited_arxiv_id":null,"evidence_quote":"Supplies the CyCTR few-shot segmentation baseline that transfers poorly to MOVE, representing the image-centric approach in the benchmark tables."},{"cited_title":"Video instance segmentation","cited_arxiv_id":null,"evidence_quote":"Provides YouTube-VIS as the category-based comparison dataset in the necessity study."}],"review_version":1}