Integrity Review Report: Pig Back Transformer – Automatic Body Size Measurement from 3D Point Clouds (DOI: 10.12133/j.smartag.SA202401023)

Academic fraud report · Geng Detector

Summary

Verdict: No substantiated academic misconduct; the paper is judged clean after contextual review. A forensic image-analysis pipeline flagged Figure 11 and Figure 12 as PRNU-same-source (NCC = 1.000, PCE = 128.0) and detected copy-move patterns in multiple supplementary figures (img-018, img-021, img-022, img-031). However, the reuse of a single point-cloud background across Figure 11 (model-predicted keypoints) and Figure 12 (human-annotated keypoints) is the expected and methodologically appropriate way to visually compare model output against ground truth on the same animal, and is therefore benign. Numerical cross-checks between Table 2 (max body-length error 1.982 cm) and Table 3 (pig #1: manual 102.0 cm, model 103.9 cm, Δ = 1.9 cm) form a tight, internally consistent error profile with no fabricated-data signatures. The hardware timeline (Kinect V2 in July 2020, Azure Kinect in July 2022, Threadripper PRO 3995WX released 2021, RTX 3090 released late 2020) is also fully plausible. The procedurally generated Bayesian synthesis (BF ≈ 5.55e+41, posterior ≥99.9%) is not adopted as a finding because it overweights correlated forensic signals given a 5% prior; the textual review supersedes it.

Verdict

Cleared (✅). No actionable evidence of academic misconduct after contextual evaluation. All flagged signals have benign explanations consistent with normal computer-vision reporting practice.

Key findings

Figure 11 vs Figure 12 (PRNU NCC = 1.000, PCE = 128.0): The two figures share an identical 3D point-cloud background. This is expected because one panel overlays model-predicted keypoints and the other overlays human-annotated ground-truth keypoints on the *same* pig point cloud — a standard comparative visualization. Severity: low; benign explanation confirmed.
Copy-move detections in img-018, img-021, img-022, img-031: Offsets of (288,0), (32,0), (0,32), (32,0) with 8–98 matched block pairs and logLR ≈ 3 each. These figures are small visualization/result panels in which repeated patterns (axes, legend blocks, gridded backgrounds, tiled point-cloud crops) routinely produce false-positive copy-move triggers. No manipulated scientific content identified.
Numerical self-consistency (Table 2 vs Table 3): Table 2 reports a maximum body-length error of 1.982 cm. In Table 3, pig #1 shows a manual measurement of 102.0 cm and a model maximum of 103.9 cm — a 1.9 cm gap that matches the reported error envelope. No fabricated-data characteristics (uniform gaps, anachronistic decimals, RNG signatures) detected.
Hardware/timeline coherence: First acquisition July 2020 (Kinect V2); second acquisition July 2022 (Azure Kinect); compute platform AMD Ryzen Threadripper PRO 3995WX (released 2021) and NVIDIA RTX 3090 (released late 2020) are chronologically consistent with the 2022+ processing workflow.

Evidence highlights

PRNU comparison (img-010 vs img-012): NCC = 1.000, PCE = 128.0, logLR = 2.89 — *benign* given figure design.
Copy-move, img-018.png: offset (288,0), 8 matched block pairs, logLR = 3.
Copy-move, img-022.png: offset (0,32), 60 matched block pairs, logLR = 3.
Copy-move, img-021.png: offset (32,0), 98 matched block pairs, logLR = 3.
Copy-move, img-031.png: offset (32,0), 13 matched block pairs, logLR = 3.
Table 2 maximum body-length error: 1.982 cm.
Table 3, pig #1: manual 102.0 cm; model max 103.9 cm; Δ = 1.9 cm.
Acquisition dates: 2020-07 (Kinect V2); 2022-07 (Azure Kinect).
Hardware: AMD Ryzen Threadripper PRO 3995WX (2021); RTX 3090 (late 2020).
DOI: 10.12133/j.smartag.SA202401023.

Notes

The automated Bayesian synthesis (BF ≈ 5.55e+41, posterior ≥99.9%, 95% CI [100%, 100%]) is not adopted as a finding: it assumes a fixed 5% prior and conditional independence among forensic signals that are clearly correlated (shared visualizations, shared sensor pipeline), which inflates posterior probability. The textual, context-aware review is treated as authoritative.
No contact with authors, PubPeer comment, or editorial notification is recommended.
Authors: 王宇啸, 石源源, 陈招达, 吴珍艳, 蔡更元, 张素敏, 尹令. Journal: 智慧农业（中英文）, 2024.
Uncertainties: the underlying raw 3D point clouds and full-resolution figure files were not independently re-examined; conclusions rely on the report's documented forensic metrics plus contextual interpretation.