Integrity Review of "Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud" (CVPR 2020)

Academic fraud report · Geng Detector

Summary

Verdict: No evidence of academic misconduct was identified in this paper (DOI: 10.1109/CVPR42600.2020.00178). The review applied the six-method integrity scan (Geng framework) to text-extracted content, focusing on numerical consistency, citation timeline, and reporting transparency. Key verifications confirmed: (1) the KITTI split (3712 train + 3769 val = 7481) is arithmetically exact; (2) the claimed 3.45-point gain over STD in BEV Easy Car detection matches Table 2 (93.11 − 89.66 = 3.45); (3) the ablation honestly reports that T=3 (88.31) underperforms T=2 (88.37), with a plausible explanation offered; (4) citations and hardware (Xeon E5-1630, GTX 1070) align with the 2020 timeframe; (5) code is publicly released on GitHub. Limits: image-based forgery checks (image reuse, splicing) could not be performed because only text was available; Figures 1–4 were not pixel-inspected. This is a computer-vision paper where such image-fabrication risks are typically low. Overall confidence in the "clean" verdict is moderate-to-high for data integrity, with image-integrity assessment deferred.

Verdict

Clean (✅ No misconduct indicators identified). The text-extracted content shows consistent numerical logic, plausible citations, anachronism-free hardware descriptions, and transparent reporting of mixed ablation results. The paper is recommended as a positive example of research integrity.

Key findings

Dataset partition arithmetic verified. Section 4.1 states 7481 KITTI training samples; Section 4.4 splits this into 3712 train + 3769 val. The sum 3712 + 3769 = 7481 is exact.
Performance gain verified. Section 4.3.1 claims a +3.45 improvement over STD on Easy BEV Car. Table 2: Point-GNN = 93.11, STD = 89.66; 93.11 − 89.66 = 3.45 matches precisely.
Honest ablation reporting. Table 4 reports T=3 at 88.31, slightly below T=2 at 88.37. The authors disclose this and attribute it to training difficulty rather than hiding it.
Citation timeline consistent. Latest references (PointPillars, STD) are 2019; submission to CVPR 2020 is chronologically coherent.
Hardware realistic for era. Xeon E5-1630 and GTX 1070 were standard research configurations circa 2020.
Open-source code released at https://github.com/WeijingShi/Point-GNN, a strong positive signal for reproducibility.

Evidence highlights

Table 2 (BEV Car Easy): Point-GNN 93.11 vs. STD 89.66 → Δ = 3.45.
Table 4 ablation: T=1 / T=2 / T=3 values reported with a non-monotonic trend (88.37 → 88.31), openly discussed.
Dataset count: 3712 + 3769 = 7481 (Section 4.1 / 4.4).
DOI: 10.1109/CVPR42600.2020.00178.

Notes

Scope limitation. Image-based checks (Geng methods 1 and 3 — image reuse, splicing) were not performed because only text was provided. Figures 1–4 were not pixel-inspected. For a CV/algorithm paper, such risks are typically low, but absence of image analysis means the "clean" verdict should be read as conditional on that gap.
Confidence. High confidence in numerical and textual consistency; moderate confidence overall due to the missing image layer. No follow-up contact with the authors or reporting action is recommended. The paper may be cited as a positive integrity exemplar.