Geng integrity report: Pig Back Transformer for automatic body size measurement from 3D point clouds (DOI: 10.12133/j.smartag.SA202401023)

Academic fraud report · Geng Detector

Summary

This report evaluates the paper "Pig Back Transformer" by Wang Yuxiao et al., published in Smart Agriculture in 2024, for potential academic misconduct. The overall verdict is clear (no substantive fraud indicators). Three areas were investigated. First, visual forensic signals (sensor-noise similarity and copy-move patterns between Figure 11 and Figure 12) were flagged, but these were assessed as benign: the figures show model-generated keypoints and human-labeled keypoints on the same input point cloud, so identical backgrounds are expected. Second, numerical analyses of Tables 2-4 showed realistic error distributions; notably, the front-body-height standard deviation (2.199) far exceeds the mean (1.132), an irregularity inconsistent with fabricated data. Third, the equipment timeline is internally consistent: a Kinect V2 (512x424) acquisition in July 2020 and an Azure Kinect (1024x1024) in July 2022, with submission in January 2024. Confidence in the verdict is moderate-high, limited by reliance on a single PDF and image-forensic heuristics that can produce false positives.

Verdict

Cleared / No substantive fraud evidence found. All flagged signals have plausible benign explanations consistent with standard computer-vision methodology and the paper's described experimental design.

Key findings

Image background similarity and PRNU/copy-move matches between Figure 11 and Figure 12 are explained by both panels displaying different annotations (model-predicted vs. human-labeled keypoints) on the same input point cloud, a standard CV visualization practice.
Numerical error data in Tables 2-4 display realistic physical-noise characteristics, including uneven last-digit distribution and high variance (front-body-height SD = 2.199 vs. mean = 1.132), arguing against fabrication by random-number generation.
An anomalous 109.0 cm measurement for pig #2 in Table 3 (vs. ~100 cm in other trials) is acknowledged by the authors and attributed to extreme posture-induced outliers.
Equipment and timeline are coherent: Kinect V2 (512×424) used July 2020, Azure Kinect (1024×1024) used July 2022, manuscript received 21 January 2024.

Evidence highlights

DOI: 10.12133/j.smartag.SA202401023
Figure 11 vs. Figure 12: img-010 and img-012 show high sensor-noise (PRNU) similarity; algorithm flowcharts (img-000, img-001, etc.) contain regular copy-move match blocks—attributed to repeated axis/text elements and a unified export pipeline.
Table 2: front-body-height SD = 2.199, mean = 1.132; rear-body-height SD = 0.678.
Table 3: pig #2, trial 4 outlier at 109.0 cm amid ~100 cm trials.
Acquisition timeline: 2020-07 (Kinect V2, 512×424), 2022-07 (Azure Kinect, 1024×1024); submission 2024-01-21.

Notes

Confidence is moderate-high but not absolute: analysis was limited to the submitted PDF; image-forensic heuristics (PRNU, copy-move) are prone to false positives under batch figure export.
The authors are encouraged to release code and a de-identified subset of point clouds to improve reproducibility and public trust.
No recommendation for formal investigation is made on the basis of the current evidence.