{"id":"3fe1b6b8-cb80-40e2-a9a1-b8f2a82a52a6","arxiv_id":"2501.07087","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Videos can be squeezed to roughly 0.2% of their original data while retaining acceptable quality prediction accuracy, enabling the lightweight online VQA model MGQA.","lead":"This paper tests how aggressively a video can be spatially and temporally downsampled before video quality assessment models lose accuracy. It finds that keeping only a tiny fraction of the video, such as a few small patches from a few frames, still gives acceptable quality predictions, and introduces a lightweight online VQA model named MGQA.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 99.83% 'computational cost reduction' (contribution 3, Table IX) is an input-pixel-count ratio, not a measured cost reduction; the online-efficiency claim is unsupported and could overstate real savings.","rationale":"The reader's weakest assumption is exactly the one I find most load-bearing: the 99.83% reduction is a data-volume ratio, not a computational-cost measurement. The central claim of the paper is two-fold: (1) aggressively sampled video retains enough information for VQA, and (2) this enables an online VQA model with a 99.83% cost reduction. The accuracy part is supported by Tables II-V as an empirical observation, and the paper explicitly limits it to datasets without extreme motion, which is honest. The efficiency part, however, is the distinctive quantitative headline, and it is currently an upper bound on data reduction rather than a demonstrated cost reduction. Because MGQA uses a different backbone and an added graph module, and because preprocessing and decoding overhead are ignored, the actual computational saving could be materially smaller than 99.83%. I also noticed an internal inconsistency between the default ECO temporal sampling (20/21 frames) and the frame counts in Table IX, which reinforces the need for a re-derivation and measurement. No machine-checked proofs or shipped code are provided, so this cannot be verified from the manuscript alone. My recommendation is UNCHANGED because the concern matches the reader's conditional verdict: the sampling study stands, but the headline efficiency claim must be re-supported with real measurements before acceptance.","tokens_in":17539,"tokens_out":6759,"duration_ms":66847,"concrete_test":"Profile VSFA and MGQA end-to-end on the same hardware (same GPU/CPU, same batch size, same video files) across the six databases, measuring wall-clock time and FLOPs for the full pipeline including temporal sampling, grid patch extraction/concatenation, feature extraction, and quality regression. If MGQA's measured cost is not close to 0.17% of VSFA's on average, or if re-derived Table IX frame counts differ from the reported ones under the specified ECO setting, the 99.83% figure should be replaced with the measured efficiency number.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In good faith, the joint spatial/temporal sampling study is a reasonable empirical contribution, and the paper's own limitation paragraph correctly confines the accuracy claim to common, mostly spatially distorted databases. The load-bearing weakness is the headline efficiency claim. Section IV-C and Table IX report 'computational cost' savings of 99.76-99.93% computed solely from input tensor dimensions (e.g., VSFA 208*540*960 vs MGQA 10*160*160). This equates 'processed data' with computational cost, which is not justified: MGQA replaces ResNet50+GRU with MobileNet plus a graph module, so per-pixel FLOPs and memory traffic differ; the spatial sampling pipeline (grid patch extraction, concatenation) and video decoding add overhead independent of the reduced tensor; and for online processing, wall-clock latency or energy is the relevant quantity. The paper reports no runtime, FLOPs, or energy measurement anywhere. Additionally, Table IX's MGQA frame counts (10, 16, 22, 15, 12, 14) do not match the paper's stated default temporal sampler ECO, which is described as using 20 or 21 keyframes (Section IV-A3), so even the input-size ratio is not reproducible as written. Because the central 'online VQA' contribution is defined by this 99.83% number, the claim needs a measured efficiency comparison before it can be accepted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates how much spatial and temporal information can be removed from videos before a no-reference VQA model degrades unacceptably. It first measures VSFA and a Transformer-based variant under combinations of four spatial sampling schemes (grid-based patch extraction) and three temporal sampling strategies (TSN, TSM, ECO) on six public databases. It then proposes MGQA, a lightweight model using MobileNet and a graph network, and reports that its processed data is about 99.8% smaller than VSFA's, which is claimed as a computational cost reduction. The paper concludes that heavily squeezed video can predict original video quality and that an online VQA model is feasible.","tokens_in":17809,"tokens_out":3470,"duration_ms":33401,"significance":"If the efficiency claim is validated with actual compute measurements, the MGQA design would be a practical contribution to low-resource VQA. The systematic sampling study itself is a useful empirical reference for understanding the trade-off between input size and quality prediction accuracy, and the paper's explicit limitation statement about high-motion videos is honest. However, the headline 99.83% figure currently rests on an unvalidated proxy, and the absence of comparisons with other efficient VQA models limits the assessment of the method's relative value.","major_comments":[{"comment":"The '99.83% computational cost reduction' claim is computed as the ratio of input tensor element counts (VSFA input dimensions versus MGQA input dimensions), not as any measured computational cost. The paper reports no wall-clock time, FLOPs, or energy measurements anywhere, even though MGQA replaces ResNet50+GRU with MobileNet plus a graph module and adds the GMS patch-extraction pipeline. As written, the headline efficiency claim is an assertion about processed data volume, not about computational cost, and the paper provides no evidence that the two are interchangeable. Please provide direct efficiency measurements (latency, FLOPs, or energy) under the same hardware and protocol, or revise the claim to state precisely that what is reduced is input pixel count.","section":"Section IV-C, Table IX, and Introduction contribution 3"},{"comment":"The MGQA default temporal sampler is stated to be ECO, which the text describes as extracting M=20 segments, with a possible extra keyframe when N mod M is nonzero (i.e., 20 or 21 keyframes). However, Table IX lists MGQA input frame counts of 10, 16, 22, 15, 12, and 14 for the six databases. These numbers are inconsistent with the described ECO default, making the reported processed-data ratios non-reproducible as written. Please clarify the actual frame counts and explain how they were derived.","section":"Section IV-A3 and Table IX"},{"comment":"The sentence 'The model consistently performs well, demonstrating state-of-the-art results on the CVD2014 and LIVE-Qualcomm datasets' is not supported by any comparison with existing BVQA methods under the same training/test protocol; Tables V-VIII only compare MGQA variants against each other. The 'state-of-the-art' claim should be substantiated with a table of recent lightweight or efficient VQA methods evaluated on the same splits, or removed.","section":"Section IV-C (page 7)"},{"comment":"The claim that joint spatial and temporal sampling 'does not lead to a significant drop in performance across the datasets' is not uniformly supported by the tables. For example, on LIVE-Qualcomm, VSFA with TSM and S1 drops from PLCC 0.73 to 0.57; on LIVE-VQC, the TSM/S1 setting drops from PLCC 0.72 to 0.63. The paper's own text concedes 'a significant performance drop on the temporal distortion dominated databases.' The conclusion should be scoped accordingly, e.g., to spatial-distortion-dominated databases, and the abstract and conclusion should be amended to match.","section":"Section IV-B, Tables II and III"}],"minor_comments":[{"comment":"The phrase 'we set the sampling stepstep to' contains a typo; it should be 'step'.","section":"Section IV-A3"},{"comment":"'sptaiotemporal sampling' appears to be a typo for 'spatiotemporal sampling'.","section":"Section II-B"},{"comment":"The row labeled 'Transofrmer' should be 'Transformer'.","section":"Table VII"},{"comment":"The text 'the Grap and fully connected layers' has an incomplete word; it should likely be 'the Graph and fully connected layers'.","section":"Section IV-C"},{"comment":"The caption states that 'TSM extracts four frames ... while TSM only samples one frame from each segment,' which confuses TSN and TSM; the second occurrence should presumably refer to TSN.","section":"Figure 3 caption"},{"comment":"'Followed by our previous study [13]' should be 'Following our previous study [13]'.","section":"Section IV-D"}],"recommendation":"major_revision","confidential_remarks":"The paper's sampling study is reasonably sound, but the central efficiency contribution is presented through an input-data proxy rather than measured computational cost. I recommend requiring direct efficiency measurements before acceptance, and asking the authors to resolve the inconsistency between the stated ECO default and the frame counts in Table IX. The 'state-of-the-art' claim on CVD2014 and LIVE-Qualcomm also needs either proper comparisons or removal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on Yan et al. (2501.07087). The sampling study is a legitimate and useful empirical contribution: it systematically varies spatial and temporal sampling together on six VQA databases and shows that aggressive reduction (down to a few small patches per video) can preserve accuracy on spatially dominated databases. That is genuinely new — prior work looked at temporal alone or spatial alone. MGQA is a simple lightweight instantiation, not a breakthrough, but the empirical grid is worth having.\n\nThe weak spot is the headline efficiency claim. Contribution 3 and Table IX claim about 99.83% computational cost reduction, but the number is computed as the ratio of input tensor pixels (VSFA's full frames vs MGQA's 10–22 160×160 patches). That is not computational cost. MGQA uses MobileNet plus a graph module, not ResNet50+GRU, so per-pixel FLOPs and memory behavior are different; the patch extraction and concatenation also add overhead; for online processing what matters is wall-clock latency or energy. The paper reports no runtime, no FLOPs, no energy. The stress-test is right: as written the central 'online VQA' claim is not supported.\n\nAlso, Table IX's frame counts don't match the described ECO default. The text says ECO uses 20 or 21 fragments, but the table lists 10, 16, 22, 15, 12, 14 frames for the six databases. Either they used a different configuration or the description is wrong; either way the cost ratio isn't reproducible. Minor: Figure 3's caption confuses TSN and TSM, and there are a few typos in the ECO formula.\n\nWhat the paper does well: the performance tables support the qualitative conclusion, the paper's own limitation paragraph correctly confines the accuracy claim to common, mostly spatially distorted content and warns about high-motion video, and the study is honest about the redundancy premise being partly established in prior work [12,13,15,16]. I don't see circularity.\n\nWho benefits: practitioners designing low-cost VQA pipelines and researchers working on VQA efficiency. The paper deserves a serious referee, because the sampling grid is useful and the flaws are addressable: measure real compute, align the frame counts, and compare MGQA against at least one or two competing efficient models under the same protocol. A revision that fixes those would be solid. I'd accept it to peer review conditionally, with a request for actual efficiency measurements.","headline":"A useful empirical grid on joint spatial/temporal sampling for VQA, but the 99.83% cost-reduction headline is an input-pixel ratio, not measured compute.","tokens_in":18353,"tokens_out":2402,"would_cite":true,"duration_ms":23203,"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":"Heavily squeezed video—a few patches from a few frames—still predicts video quality on common in-the-wild databases.","keywords":["video quality assessment","no-reference VQA","spatial-temporal sampling","temporal keyframe sampling","grid mini-patch sampling","lightweight video quality model","MGQA","graph-based quality regression"],"falsifier":"Run VSFA and MGQA on the same hardware with the exact input sizes of Table IX and measure end-to-end latency and FLOPs; if the runtime reduction is far smaller than the $99.83\\%$ data reduction, the paper's computational-cost claim fails as stated. Alternatively, evaluate the trained MGQA on a high-motion video database with rapidly moving objects or frequent scene changes; if PLCC/SRCC drop below acceptable levels, the claim that heavily squeezed video predicts original quality does not generalize to dynamic content.","tokens_in":17332,"feed_emoji":"🎞️","tokens_out":8062,"duration_ms":71495,"temperature":0.7,"pith_summary":"This paper asks how little of a video a quality-assessment model actually needs to see. It argues that videos are so redundant in space and time that a heavily squeezed video—a few small patches from a few sampled frames—can predict the perceptual quality of the original video with acceptable accuracy on existing in-the-wild databases. Across six databases, jointly sampling temporal keyframes and spatial grid patches keeps performance close to feeding the full video, and the resulting MGQA model cuts the amount of processed pixel data by an average of $99.83\\%$ relative to the VSFA baseline. If true, online and mobile video-quality monitoring could run on roughly $1/500$ of the pixel data without needing full-resolution video as input.","feed_headline":"Videos can be quality-scored from just 0.2% of the data","feed_subtitle":"A model fed tiny sampled patches and frames matches full-video quality scores across six in-the-wild databases.","key_machinery":"The mechanism is a two-stage spatio-temporal sampling pipeline. Temporally, the video is divided into segments and keyframes are drawn by one of three strategies (TSN, TSM, or ECO); spatially, each keyframe is divided into uniform grids and small patches (e.g., $32\\times32$ or $64\\times64$) are randomly sampled from every grid and stitched into a fragment, following the GMS method. This produces stacked spatio-temporal blocks that retain enough local distortion evidence for quality regression while discarding most pixels. The MGQA model then pairs MobileNet as a lightweight spatial distortion capturer with a graph-based temporal fusion module and fully connected global regression.","core_discovery":"The central claim is that the heavily squeezed video can be used to predict the quality of the original video: after aggressive temporal sampling (TSN, TSM, or ECO) and spatial grid-patch sampling (GMS), a handful of small fragments—for example $10 \\times 160 \\times 160$ pixel patches on KoNViD-1k instead of $208 \\times 540 \\times 960$—suffices to keep correlation with human scores close to the full-input model. The paper shows this for VSFA and a transformer variant across six databases, and then instantiates an online model, MGQA, whose MobileNet spatial extractor and graph-network temporal fusion achieve PLCC/SRCC values such as $0.87/0.89$ on CVD2014 and $0.83/0.86$ on LIVE-Qualcomm while processing about $0.17\\%$ of the original pixel data on average. The authors read this as demonstrating the feasibility of online VQA through joint sampling and a deliberately simple architecture.","pith_inferences":["The paper's 'computational cost' is measured as processed data volume, not runtime; an equal-hardware latency and energy benchmark is the natural next experiment and would either support or qualify the $99.83\\%$ claim.","The orderless robustness suggests that much of VQA's difficulty is spatial; an implied testable extension is to combine aggressive sampling with a motion-aware light module for high-motion and HDR content, which the paper explicitly flags as an open problem.","The large backbone gap implies that data reduction alone does not determine efficiency; a future study could treat sampling density and backbone width as a joint trade-off curve rather than fixing one.","The graph regressor's advantage over recurrent and transformer fusion on tiny inputs hints that relational pooling over patch positions is well matched to fragment-based input, a hypothesis the paper does not directly test."],"forward_implications":["Video quality monitoring can be shifted to tiny inputs: on the six tested databases, roughly $0.17\\%$ of the pixel data suffices to keep correlation within a small margin of the full-input VSFA baseline.","Aggressive joint sampling is safest where spatial distortion dominates (CVD2014, KoNViD-1k, LSVQ); on temporal-distortion-heavy content (LIVE-VQC, LIVE-Qualcomm) the paper reports larger drops, so sampling density should be content-adaptive.","The order of sampled frames barely matters, so online systems can process short, shuffled fragments rather than long ordered sequences.","A graph-based temporal fusion regressor is the key to MGQA's accuracy: replacing it with GRU, LSTM, transformer, or plain FC degrades LIVE-VQC performance substantially in the paper's ablation.","Lightweight spatial backbones are not interchangeable: among MobileNet, FasterNet, EfficientNet, ShuffleNet, and MobileOne, only MobileNet keeps PLCC above 0.65 on LIVE-VQC, which points to backbone choice as a major efficiency-accuracy lever."],"supporting_citations":[{"why":"It supplies the stable VQA baseline (VSFA, ResNet50 plus GRU) whose input sizes and processed-data cost are the comparison target for the 99.83% reduction claim.","marker":"[14]"},{"why":"It provides the GMS fragment sampling method used as the default spatial sampling strategy.","marker":"[15]"},{"why":"It provides one of the three temporal sampling strategies, TSN, whose keyframe extraction is tested in the joint sampling study.","marker":"[31]"},{"why":"It provides the TSM temporal sampling strategy used in the experiments.","marker":"[30]"},{"why":"It provides the ECO sparse temporal sampling strategy, which is the default for MGQA.","marker":"[33]"},{"why":"It is the prior study establishing that sparse or shuffled-frame input retains VQA performance, the starting point extended here to joint spatial-temporal sampling.","marker":"[13]"},{"why":"It is the prior finding that few sparsely sampled frames predict free-viewpoint video quality acceptably, motivating temporal sampling.","marker":"[12]"},{"why":"It provides MobileNet, the lightweight spatial feature extractor that gives MGQA its accuracy-efficiency profile.","marker":"[53]"},{"why":"It provides the graph network used as temporal fusion/regressor, and the ablation shows it outperforms GRU, LSTM, transformer, and FC on LIVE-VQC.","marker":"[54]"}],"fun_headline_variants":["Video quality scores from only 0.2% of data","Aggressive sampling preserves VQA accuracy","Online VQA with sparse spatial-temporal input","Minimal video data still predicts quality well","Joint spatial-temporal sampling for efficient VQA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that reducing the amount of input data by $99.83\\%$ reduces computational cost by roughly the same proportion, since no wall-clock time, FLOPs, or energy is measured and the architecture changes from ResNet50 plus GRU to MobileNet plus Graph.","fun_headline_variants_meta":{"raw":{"variants":["Video quality scores from only 0.2% of data","Aggressive sampling preserves VQA accuracy","Online VQA with sparse spatial-temporal input","Minimal video data still predicts quality well","Joint spatial-temporal sampling for efficient VQA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000282,"raw_usage":{"total_tokens":1698,"prompt_tokens":1003,"completion_tokens":695,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":626}},"tokens_in":619,"tokens_out":695,"duration_ms":7187,"temperature":1.0,"reasoning_tokens":626,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:49:27.455008+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run VSFA and MGQA on the same hardware with the exact input sizes of Table IX and measure end-to-end latency and FLOPs; if the runtime reduction is far smaller than the $99.83\\%$ data reduction, the paper's computational-cost claim fails as stated. Alternatively, evaluate the trained MGQA on a high-motion video database with rapidly moving objects or frequent scene changes; if PLCC/SRCC drop below acceptable levels, the claim that heavily squeezed video predicts original quality does not generalize to dynamic content.","supporting_citations":[{"cited_title":"Quality assessment of in-the-wild videos,","cited_arxiv_id":null,"evidence_quote":"It supplies the stable VQA baseline (VSFA, ResNet50 plus GRU) whose input sizes and processed-data cost are the comparison target for the 99.83% reduction claim."},{"cited_title":"FAST-VQA: Efficient end-to-end video quality assessment with fragment sampling,","cited_arxiv_id":null,"evidence_quote":"It provides the GMS fragment sampling method used as the default spatial sampling strategy."},{"cited_title":"Temporal segment networks for action recognition in videos,","cited_arxiv_id":null,"evidence_quote":"It provides one of the three temporal sampling strategies, TSN, whose keyframe extraction is tested in the joint sampling study."},{"cited_title":"TSM: Temporal shift module for efficient video understanding,","cited_arxiv_id":null,"evidence_quote":"It provides the TSM temporal sampling strategy used in the experiments."},{"cited_title":"ECO: Efficient convolutional network for online video understanding,","cited_arxiv_id":null,"evidence_quote":"It provides the ECO sparse temporal sampling strategy, which is the default for MGQA."},{"cited_title":"Study of spatio-temporal modeling in video quality assessment,","cited_arxiv_id":null,"evidence_quote":"It is the prior study establishing that sparse or shuffled-frame input retains VQA performance, the starting point extended here to joint spatial-temporal sampling."},{"cited_title":"GraphIQA: Learning distortion graph representations for blind image quality assessment,","cited_arxiv_id":null,"evidence_quote":"It provides the graph network used as temporal fusion/regressor, and the ablation shows it outperforms GRU, LSTM, transformer, and FC on LIVE-VQC."}],"review_version":1}