Integrity Review Report: MMSCN: Mask Based Multi-interval Spatiotemporal Convolution Network for Drop Detection (Geng Fraud Detection, ID geng_geng_6aedab2b386626.71195423)

Academic fraud report · Geng Detector

Summary

This report assesses the paper "MMSCN: Mask Based Multi-interval Spatiotemporal Convolution Network for Drop Detection" (Dai, Tian, Jiang, Lyu; IPPR 2025; DOI 10.1109/IPPR66507.2025.11198569) and concludes it is highly suspicious (🟠). Confirmed textual defects include: (1) multiple Acc/Recall/F1 triplets in Table 2 that violate the mathematical constraint that F1 must lie between precision and recall (e.g., TSN: Acc 0.33, Recall 0.10, F1 0.33, where F1's upper bound is ≈0.18); (4) three mutually contradictory definitions of the multi-interval experiment (even intervals 2–16, odd intervals 3/5/7/9, and 1–15); and (7) misaligned citations (TSA-Net and I3D cited with wrong references). Additional suspected issues: a stated ablation improvement of 8.0% in J versus an actual table difference of 7.1%; copy-move forensics in Figure 1 (750 matched block pairs at offset (32,0)) and Figure 4, suggesting duplicated static frames inconsistent with the claimed real driving scenarios; these remain unconfirmed pending original materials. A Bayesian synthesis flags ≥99.9% posterior suspicion, though this assumes conditional independence and a fixed 5% prior. Final determination requires institutional investigation.

Verdict

Highly suspicious (🟠). Multiple independent, verifiable internal inconsistencies were confirmed, including mathematically impossible F1 values and contradictory experimental setups. Image-forensics leads remain unconfirmed suspicions. Final determination requires formal investigation; AI-assisted report, for discussion only.

Key findings

Confirmed — Table 2 F1 values mathematically impossible (🔴): Several Acc/Recall/F1 triplets violate F1 = 2PR/(P+R) bounds: TSN (video frames) Acc 0.33, Recall 0.10, F1 0.33 (F1 upper bound ≈0.18); TSM Recall 0.12, F1 0.09 (F1 cannot be below recall); Slowfast Recall 0.12, F1 0.33 (bound ≈0.21); VideoMAEv2 Recall 0.46, F1 0.22. No benign explanation (rounding, SEM/SD confusion) can produce F1 beyond its mathematical maximum.
Confirmed — Contradictory multi-interval definitions (🟠): Section C defines intervals 2,4,6,8,10,12,14,16 (⟨D2…D16⟩); Section F claims single intervals (3,5,7,9) vs multi-interval 1–15. Three inconsistent definitions of the same core module; the ablation supporting +5.1% acc, +12.29% recall, +6.67% F1 is shown only as curves (Figure 6) with no numerical table.
Confirmed — Citation misalignment (🟡): Introduction cites "TSA-Net[10]" but [10] is Cheng et al. 2023 "Segment and track anything"; "I3D[7]" but [7] is Ren et al. PoseTransformer3D; I3D (Carreira & Zisserman) absent from the reference list.
Suspected — Ablation discrepancy (🟡): Table 3 text claims J improvement of 8.0%, but table values give 97.1 − 90.0 = 7.1% (not explainable by rounding). F (13.8%) and J&F (10.9%) match.
Suspected — Figure 1 frame reuse (🔴 severity, ⚠️ status): Frames 16/24/32/48 in the "Video" row are near-identical in background, cargo, and lighting despite claimed real driving scenarios; forensic analysis found 750 exact-matching block pairs (copy-move, regular offset). Benign explanations unsupported in the text; original video needed.
Suspected — Figure 4 dataset sampling duplication (🟠 severity, ⚠️ status): Highly similar images in the Straight scene row; copy-move matches at offset (32,0) (42 and 50 block pairs) and noise-variance CV=1.475 suggesting possible splicing. Dataset claimed as "the first dataset dedicated to object-drop anomalies" (54 videos, 2778 frames).
Not upheld: MMSCN's 96%/98%/99% combination judged "statistically implausible" lacks mathematical impossibility support; retained only as an unnatural-perfection concern. Benford/digit tests excluded (data prerequisites unmet).

Evidence highlights

Table 2 values verified verbatim against the paper; F1 impossibilities are hard formula-level contradictions.
Copy-move forensics: img-005.jpg 750 matched pairs at offset (32,0); img-002.jpg 42 pairs; img-004.jpg 50 pairs.
All quoted figures (96/98/99, 91, Table 3 values) confirmed present in the source text.
Automated Bayesian synthesis: posterior ≥99.9% (5% prior), BF = 2.35e+7; ⚠️ this synthesis assumes conditional independence and may overstate correlated evidence.

Notes

Report is AI-assisted and for academic discussion only; formal findings require institutional (East China Normal University) and venue (IPPR/IEEE) investigation.
Image-forensics findings (Figures 1 and 4) have no in-text benign support but remain unverified suspicions pending original video material.
Recommended actions: request raw logs and per-frame predictions for Tables 2–3; explain interval contradictions and provide Figure 6 numerics; publish/verify the dataset; raise on PubPeer; report to editors and institutional committee.
DOI: 10.1109/IPPR66507.2025.11198569