{"id":"59d796e7-81d9-4354-81f6-7e0821f74fff","arxiv_id":"2501.06591","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"PoseLift is a real-world privacy-preserving pose dataset of 155 retail CCTV videos, and the best benchmarked model, STG-NF, reaches 67.46% AUC-ROC for shoplifting detection.","lead":"This paper introduces PoseLift, a privacy-preserving dataset of human body poses extracted from real CCTV footage of normal shopping and shoplifting in a retail store. The authors benchmark three pose-based anomaly detection models and report moderate accuracy, offering a new testbed for privacy-friendly retail security.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'high detection accuracy' claim rests on 67.46% AUC-ROC and 0.39 EER from Table 3, with no chance baseline or confidence intervals; with only 19 normal and 43 anomalous test scenarios, frame-level metrics give no statistical support for above-chance detection.","rationale":"The dataset itself may be a valuable contribution: real-world, multi-camera, pose-only shoplifting data is scarce, and the pipeline description is concrete. My concern is not that the pose extraction is obviously broken; it is that the benchmark does not currently establish the paper's central performance claim. Even if the poses preserve behavior, the reported numbers are modest and statistically uncharacterized. The proposed scenario-bootstrap check is the minimal experiment that would resolve this. If the confidence intervals exclude chance and are reasonably tight, the claim can stand; if not, the abstract and conclusion need to be softened to 'pose-based detection is feasible but accuracy remains modest' or the evaluation needs stronger baselines. Thus I would keep the reader's CONDITIONAL verdict rather than accepting the paper as-is.","tokens_in":10786,"tokens_out":7386,"duration_ms":73893,"concrete_test":"Re-run the STG-NF evaluation on PoseLift and report scenario-level bootstrap 95% confidence intervals for AUC-ROC and AUC-PR, resampling the 43 shoplifting and 19 normal test scenarios with replacement, plus a random-score baseline. If the AUC-ROC confidence interval includes 0.50, or the AUC-PR interval overlaps the positive-frame prevalence of about 0.40, the 'high detection accuracy' claim is not supported and the abstract/conclusion must be softened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline contribution is that pose-based models 'achieve high detection accuracy' on PoseLift. The strongest model, STG-NF, reaches AUC-ROC 67.46%, AUC-PR 84.06%, and EER 0.39 (Table 3). These numbers do not by themselves establish high accuracy: at the EER operating point, roughly 39% of frames are misclassified, and AUC-ROC is only modestly above the 50% chance level. More importantly, the test set contains only 19 normal scenarios and 43 shoplifting scenarios (Table 2), while the reported metrics are computed over 2,221 normal and 1,500 anomalous frames that are heavily autocorrelated within each tracked person. No confidence intervals, per-scenario scores, or a random baseline are reported. Because the effective independent sample size is around 62 scenarios, the claim that pose sequences preserve enough behavioral signal to detect shoplifting could rest on a few favorable tracks. The dataset may be useful regardless, but the benchmark does not currently support the abstract's accuracy claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces PoseLift, a privacy-preserving dataset of human pose sequences extracted from real CCTV footage of a retail store, with frame-level labels for normal shopping and shoplifting. It frames shoplifting detection as unsupervised pose-based anomaly detection and benchmarks three existing models (STG-NF, TSGAD, GEPC) with AUC-ROC, AUC-PR, and EER. The best result is STG-NF with AUC-ROC 67.46%, AUC-PR 84.06%, and EER 0.39 on a test set of 2,221 normal and 1,500 anomalous frames from 19 normal and 43 shoplifting scenarios.","tokens_in":11132,"tokens_out":5349,"duration_ms":50784,"significance":"If the claims are supported, PoseLift would be a useful community resource because it is one of the first real-world retail shoplifting datasets in pose form, with six camera views and de-identified annotations, and it directly addresses privacy concerns that limit raw-video datasets. The benchmark is a reasonable starting point: the models are from prior work, the training/test separation is described, and the authors include a failure-case analysis. The dataset contribution is therefore credible, but the paper's headline accuracy claim currently exceeds what the evidence shows.","major_comments":[{"comment":"The abstract states that pose-based approaches “achieve high detection accuracy,” but the best measured results are AUC-ROC 67.46% and EER 0.39. At the EER operating point, 39% of frames are misclassified, and the AUC-ROC is only modestly above the 50% chance level, with no chance-level baseline or confidence interval reported. Please either soften the claim to “feasibility” or support “high accuracy” with per-scenario results and statistical intervals.","section":"Abstract; Sec. 6, Table 3"},{"comment":"The evaluation uses frame-level metrics over 2,221 normal and 1,500 anomalous frames, but these frames are strongly autocorrelated within the 19 normal and 43 shoplifting scenarios. The effective independent sample size is therefore dozens, not thousands. Reporting AUC-ROC, AUC-PR, and EER without bootstrap confidence intervals, multiple seeds, or scenario-level aggregation makes it impossible to determine whether the best model is above chance. Please add a random baseline, per-scenario scores, and confidence intervals.","section":"Sec. 6, Table 2"},{"comment":"The pose extraction pipeline is a central assumption of the benchmark: YOLOv8 plus ByteTrack plus HRNet with linear interpolation, an 8-frame smoothing window, and per-camera region-of-interest masks. The paper provides no validation that this pipeline preserves the concealment motion at 15 fps under high-angle views and shelf occlusions. Please report pose quality statistics, example tracking failures, or an ablation of the smoothing and interpolation choices; otherwise the benchmark result may be an artifact of pose noise.","section":"Sec. 3.2"}],"minor_comments":[{"comment":"The statement that AUC-ROC does not take the false negative rate into account is incorrect; the ROC curve uses TPR = 1 - FNR, so FNR is accounted for by TPR.","section":"Sec. 5.1"},{"comment":"There is a typo: “all models, except STG-NG, used the Adam optimizer” should read “except STG-NF,” and “Adamx” should likely be “AdamW.”","section":"Sec. 6"},{"comment":"The UCF-Crime row lists a total length of 460,800 seconds and a dash for camera count; since UCF-Crime is an online video collection, it would be helpful to clarify that its length and camera count are not directly comparable to store-captured datasets.","section":"Sec. 4, Table 1"},{"comment":"The dataset release URL is given, but there is no license, documentation of annotation redundancy, or inter-annotator agreement measure; please specify the annotation quality assurance process.","section":"Sec. 3.2"},{"comment":"The “normality score” used in the false-positive example is not defined in Section 5; please define it or refer to the relevant model output.","section":"Fig. 7"}],"recommendation":"major_revision","confidential_remarks":"The paper would be acceptable if revised to match the evidence: the dataset is valuable despite the modest benchmark numbers. I would not require new data collection; scenario-level evaluation, confidence intervals, a chance baseline, and recalibrated language for the accuracy claim could address the main concerns. Note that one benchmarked model (TSGAD) is from the authors' own group, but it does not win, so I do not see a circularity issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The dataset is the real contribution here. PoseLift is genuinely new: real CCTV footage from a single retail store, converted to pose sequences with person tracks, spanning six camera views, and explicitly privacy-preserving. That combination does not exist in the prior datasets they compare against—staged actor sets and raw UCF-Crime videos. The benchmark itself is honest in one important way: the best model, STG-NF, is external, and their own TSGAD scores lower. No one is cherry-picking a self-owned method to win.\n\nThe soft spots are real and specific. The abstract says pose-based approaches achieve \"high detection accuracy,\" but the numbers in Table 3 are moderate at best: STG-NF gets 67.46% AUC-ROC and 0.39 EER. At the EER operating point, roughly 39% of frames are misclassified. That is not high accuracy, and the claim should be toned down.\n\nThe evaluation also lacks statistical support. No chance-level baseline, no confidence intervals, no per-scenario breakdown. The test set is 19 normal and 43 shoplifting scenarios, while the frame-level metrics are computed over heavily autocorrelated frames within each track. The effective independent sample size is around 62, so the observed separation could be driven by a few favorable tracks. Per-scenario scores and a random-prediction baseline are needed before the benchmark supports any conclusion about pose-based detection.\n\nThe pose extraction pipeline is described clearly enough to be credible, but no code or exact preprocessing details are provided, and the availability of the dataset cannot be verified from the GitHub link as of the reading. The test-set composition (few normal scenarios, more shoplifting scenarios) is acknowledged, so that is not a hidden flaw, but the paper calls it \"balanced\" while the scenario counts are actually imbalanced.\n\nThese are fixable issues. The dataset fills a real gap, and the paper is worth a serious referee. A referee should ask for a corrected abstract, per-scenario evaluation, a chance baseline, and either code or a clear preprocessing specification. If those are addressed, the benchmark conclusions will actually be supported.\n\nFor what it is worth, the citation pattern is fine. Self-citations are to relevant prior pose-anomaly work, and the authors are not leaning on their own papers to prop up the result.\n\nRecommendation: send it to peer review, but treat the accuracy claim as the main thing that needs to change.","headline":"PoseLift is a genuinely useful privacy-preserving real-world shoplifting pose dataset, but the benchmark results don't support the 'high accuracy' claim and need a statistically honest evaluation.","tokens_in":11565,"tokens_out":1655,"would_cite":true,"duration_ms":17028,"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 paper introduces PoseLift, a real-store CCTV dataset of anonymized pose sequences for shoplifting, and shows that pose-based unsupervised anomaly detectors can flag genuine concealment behavior, with the best model reaching 67.46%…","keywords":["shoplifting detection","pose-based anomaly detection","privacy-preserving dataset","human pose estimation","unsupervised anomaly detection","retail surveillance","video anomaly detection","PoseLift"],"falsifier":"Run the best STG-NF model trained on PoseLift's training split on a second, independently collected set of shoplifting videos from the same store; if its AUC-ROC drops to near chance while humans viewing the same pose sequences can still identify the concealment frames, the dataset does not preserve the behavioral signal the paper claims.","tokens_in":10587,"feed_emoji":"🛒","tokens_out":7209,"duration_ms":56330,"temperature":0.7,"pith_summary":"The paper introduces PoseLift, a privacy-preserving dataset of 155 real-world retail videos converted into anonymized pose sequences (bounding boxes, person IDs, 17-keypoint skeletons) from six camera views, including 43 genuine shoplifting incidents. It argues that framing shoplifting detection as unsupervised anomaly detection on pose data is viable: models trained only on normal shopping frames can flag the anomalous concealment motion. The benchmark reports STG-NF achieving 67.46% AUC-ROC, 84.06% AUC-PR, and 0.39 EER on a test set of 1,500 anomalous and 2,221 normal frames. If accepted, the dataset offers a path to retail security analytics that keep behavioral signal while removing identifiable video, and gives researchers a public benchmark for a task previously dominated by staged or online-sourced video.","feed_headline":"Pose data alone flags real shoplifting at 67% AUC-ROC","feed_subtitle":"Privacy-safe skeleton tracks from store CCTV train unsupervised theft detectors.","key_machinery":"The load-bearing object is the PoseLift extraction pipeline: YOLOv8 detects people, ByteTrack assigns persistent person IDs, HRNet produces COCO17 2D keypoints, linear interpolation fills missing poses, and an 8-frame window smooths the tracks. Frame-level annotations (0/1) mark shoplifting as any frame with item concealment into pockets, bags, or clothing. The evaluation machinery is the unsupervised training/test split, with 53,353 normal frames for training and a test set balanced between 1,500 anomalous and 2,221 normal frames, scored by AUC-ROC, AUC-PR, and EER.","core_discovery":"PoseLift is among the first shoplifting datasets drawn from real store CCTV rather than staged actors, and it is the first to release pose sequences instead of raw pixels. The paper's central claim is that this abstract body-keypoint representation preserves enough of the concealing motion that unsupervised pose-based anomaly detectors, trained only on normal shopping behavior, can separate shoplifting from routine activity. Evidence is the benchmark of three models: STG-NF (normalizing flows) reaches 67.46% AUC-ROC and 84.06% AUC-PR; TSGAD (variational autoencoder, pose branch) reaches 63.35% AUC-ROC; GEPC (graph embedded pose clustering) reaches 60.61% AUC-ROC, all with EER near 0.38–0.41. The paper also claims the dataset addresses privacy and bias by design, and documents a false-positive case where normal actions such as phone use from a pocket are scored as anomalous.","pith_inferences":["A natural next test is to run the same models on raw video from the same store under identical train/test splits; if raw-video models substantially exceed 67.46% AUC-ROC, then pose extraction is discarding behavior the benchmark cannot see.","The false-positive example hints that pocket-related normal actions such as phone use resemble concealment; an extension would add hand-object interaction cues or trajectory-level context to reduce this confusion.","Because training uses only normal frames, PoseLift could support continual adaptation: a store could retrain on its own normal footage and deploy without any theft labels.","The imbalance between 43 shoplifting scenarios and 19 normal test scenarios means the reported EER should be read with care; a larger normal test set would tighten the estimate."],"forward_implications":["If pose-based unsupervised detection works on PoseLift, retail security systems can analyze body-keypoint streams in real time without storing or transmitting identifiable video.","Shoplifting detection can be treated as an anomaly detection problem: a model trained on a store's normal shopping patterns can flag unusual concealment behavior without labeled theft examples.","The benchmark establishes STG-NF as the reference baseline for future work on PoseLift.","Because the dataset contains six camera views and multiple concealment styles, models trained on it may generalize across aisle layouts better than single-view staged datasets.","PoseLift's public release gives the research community a way to compare methods on genuine shoplifting data while avoiding the privacy restrictions that block raw video sharing."],"supporting_citations":[{"why":"Supplies STG-NF, the normalizing-flow pose anomaly detector that achieves the best reported results on PoseLift.","marker":"[13]"},{"why":"Supplies GEPC, the graph-embedded pose clustering baseline used for comparison.","marker":"[22]"},{"why":"Supplies TSGAD, the pose-branch variational autoencoder baseline used for comparison.","marker":"[26]"},{"why":"Provides the surveillance pose-extraction approach the authors adapt to build PoseLift annotations.","marker":"[29]"},{"why":"YOLOv8 is the person detector used to localize individuals in each frame.","marker":"[16]"},{"why":"ByteTrack assigns persistent person IDs across frames for temporal tracking.","marker":"[37]"},{"why":"HRNet extracts the 2D skeleton keypoints that form the dataset's pose sequences.","marker":"[34]"},{"why":"UCF-Crime is the existing real-world anomaly video collection used as the main comparison for real data.","marker":"[33]"},{"why":"Staged shopping-mall dataset used to contrast PoseLift's real-store collection.","marker":"[5]"},{"why":"Staged 900-video shoplifting dataset used to contrast PoseLift's real-world, multi-view design.","marker":"[23]"}],"fun_headline_variants":["PoseLift: real CCTV shoplifting data, privacy-safe, 67% AUC","Real shoplifters caught by pose-only AI: 67% AUC","Pose-only anomaly detection hits 67% AUC on real theft","First real-world pose shoplifting set: 67% AUC","PoseLift: real theft poses, privacy-safe, unsup 67% AUC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the automated pose-extraction pipeline—YOLOv8, ByteTrack, HRNet keypoints, interpolation, and the 8-frame smoothing window—keeps the behavioral signal of shoplifting intact despite high-angle views, shelf occlusions, and 15 fps capture; if the pose tracks become too noisy or lose the concealment motion, the benchmark cannot support conclusions about pose-based shoplifting detection.","fun_headline_variants_meta":{"raw":{"variants":["PoseLift: real CCTV shoplifting data, privacy-safe, 67% AUC","Real shoplifters caught by pose-only AI: 67% AUC","Pose-only anomaly detection hits 67% AUC on real theft","First real-world pose shoplifting set: 67% AUC","PoseLift: real theft poses, privacy-safe, unsup 67% AUC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000877,"raw_usage":{"total_tokens":3797,"prompt_tokens":956,"completion_tokens":2841,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":2740}},"tokens_in":572,"tokens_out":2841,"duration_ms":19384,"temperature":1.0,"reasoning_tokens":2740,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:56:17.195447+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the best STG-NF model trained on PoseLift's training split on a second, independently collected set of shoplifting videos from the same store; if its AUC-ROC drops to near chance while humans viewing the same pose sequences can still identify the concealment frames, the dataset does not preserve the behavioral signal the paper claims.","supporting_citations":[{"cited_title":"Normalizing flows for human pose anomaly detection","cited_arxiv_id":null,"evidence_quote":"Supplies STG-NF, the normalizing-flow pose anomaly detector that achieves the best reported results on PoseLift."},{"cited_title":"Graph embedded pose clustering for anomaly detection","cited_arxiv_id":null,"evidence_quote":"Supplies GEPC, the graph-embedded pose clustering baseline used for comparison."},{"cited_title":"An exploratory study on human-centric video anomaly detection through variational autoencoders and trajectory prediction","cited_arxiv_id":null,"evidence_quote":"Supplies TSGAD, the pose-branch variational autoencoder baseline used for comparison."},{"cited_title":"Ancilia: Scalable intelligent video surveillance for the artificial intelligence of things","cited_arxiv_id":null,"evidence_quote":"Provides the surveillance pose-extraction approach the authors adapt to build PoseLift annotations."},{"cited_title":"Bytetrack: Multi-object tracking by associating every detection box","cited_arxiv_id":null,"evidence_quote":"ByteTrack assigns persistent person IDs across frames for temporal tracking."},{"cited_title":"Real-world anomaly detection in surveillance videos","cited_arxiv_id":null,"evidence_quote":"UCF-Crime is the existing real-world anomaly video collection used as the main comparison for real data."},{"cited_title":"Expert video-surveillance system for real-time detection of suspicious behaviors in shopping malls","cited_arxiv_id":null,"evidence_quote":"Staged shopping-mall dataset used to contrast PoseLift's real-store collection."},{"cited_title":"Shoplifting detection using hybrid neural network cnn-bilsmt and development of benchmark dataset","cited_arxiv_id":null,"evidence_quote":"Staged 900-video shoplifting dataset used to contrast PoseLift's real-world, multi-view design."}],"review_version":1}