Integrity Review Report: GoMAvatar: Efficient Animatable Human Modeling from Monocular Video Using Gaussians-on-Mesh

Academic fraud report · Geng Detector

Summary

Verdict: No evidence of academic fraud detected. The review examined the CVPR 2024 paper by Jing Wen et al. (UIUC) across multiple consistency dimensions. Cross-table data verification showed perfect mathematical self-consistency between Table 1 main results (PSNR 30.37, SSIM 0.9689, LPIPS* 32.53), Table 4 ablation (PSNR 30.36, SSIM 0.9690, LPIPS* 33.28), and Table 5 subdivisions (30.36/0.9690/33.28 without; 30.37/0.9689/32.53 with subdivision), reflecting careful separation of full vs. baseline configurations. Mathematical logic verification confirmed the claimed 43 FPS rendering speed matches the reported 23.2 ms inference time (1000/23.2 ≈ 43.1 FPS). The authors honestly acknowledged a slightly worse Chamfer Distance (2.8364) versus HumanNeRF and MonoHuman, attributing it to 3D Gaussian thickness rather than concealing it. Reference timeline is consistent with a 2024 CVPR submission. Limitations: the review was text-based only; no pixel-level image forensics was performed.

Verdict

No fraud indicators detected (✅ Clean). The paper passes cross-table consistency, mathematical logic, methodological transparency, and timeline plausibility checks without anomalies.

Key findings

Cross-table numerical self-consistency: Table 1 (30.37 / 0.9689 / 32.53), Table 4 (30.36 / 0.9690 / 33.28), and Table 5 ablations are internally consistent and correctly distinguish the subdivided main model from the non-subdivided ablation baseline.
FPS-to-latency arithmetic validated: Stated 43 FPS corresponds to the reported 23.2 ms/frame (1000 ÷ 23.2 ≈ 43.1 FPS).
Honest disclosure of weaker metrics: Authors report Chamfer Distance 2.8364, slightly worse than HumanNeRF and MonoHuman, and explain it as a known artifact of 3D Gaussian surface thickness rather than hiding the result.
Timeline consistency: Citations to ICLR 2024 and 3DV 2024 are plausible given a CVPR 2024 submission/review cycle in late 2023 / early 2024.

Evidence highlights

Table 1, Table 4, Table 5 numeric alignment (PSNR / SSIM / LPIPS*): identical metric values appear in the configuration where they should, and differ exactly where the subdivision module is toggled.
Section 4.2 explicit acknowledgment of inferior Chamfer Distance with a physical explanation (Gaussian splatting mask thickness along surface normals).
Section 4 rendering-speed claim of 43 FPS numerically matches Table 1's 23.2 ms inference time.
DOI: 10.1109/CVPR52733.2024.00201.

Notes

Review was performed on text-extracted content only; raw high-resolution figures were not available for pixel-level forensics (e.g., splicing, copy-move, JPEG re-encoding checks).
All quantitative checks above rely solely on the figures as printed; any downstream PDF re-typesetting errors cannot be ruled out without source data or code.
A passing integrity review does not guarantee correctness of the science itself; it only indicates no fraud indicators were observed within the scope of analysis.