Integrity Review: Dome-DETR: DETR with Density-Oriented Feature-Query Manipulation for Efficient Tiny Object Detection

Academic fraud report · Geng Detector

Summary

Verdict: No substantive integrity issues identified (clean). The review examined arithmetic consistency between the abstract and result tables, ablation-study coherence, parameter/complexity accounting, figure reuse (constrained by text-only medium), and citation timeline. Key cross-checks: the abstract claim of +3.3 AP on AI-TOD-V2 (34.6 minus 31.3) and +2.5 AP on VisDrone (39.0 minus 36.5) reproduce exactly; the very-tiny-object sensitivity in Table 4 aligns with the textual claim; parameter and GFLOPs arithmetic (5.88% increase, +31.2 GFLOPs) is consistent; and the YOLOv12 (Feb 2025) citation is temporally plausible for an MM '25 submission. Confidence is high for data and logic dimensions but moderate-to-limited for image-level forgery, since only textual content was available and pixel-level PS/duplication analysis was not possible. No follow-up action recommended based on current evidence.

Verdict

No integrity concerns detected. The paper's numerical claims are internally consistent across abstract, tables, appendices, and efficiency analyses. Image-level verification was not possible because only text was provided.

Key findings

Abstract-vs-table arithmetic matches exactly (+3.3 AP on AI-TOD-V2, +2.5 AP on VisDrone).
Ablation sensitivity results in Table 4 align with the textual description (smaller H/W raises $AP_{vt}$ while lowering overall AP).
Parameter and GFLOPs accounting is arithmetically correct (+5.88% params, +31.2 GFLOPs vs. D-FINE-L).
Reference to YOLOv12 (ArXiv, Feb 2025) is chronologically plausible for an MM '25 submission.
Image-level reuse and pixel-level forensics were not assessable from the provided medium.

Evidence highlights

Table 1: Dome-DETR-L 34.6 vs. D-FINE-L 31.3 → Δ = 3.3 AP on AI-TOD-V2 (matches abstract).
Table 2: Dome-DETR-L 39.0 vs. D-FINE-L 36.5 → Δ = 2.5 AP on VisDrone (matches abstract).
Table 4: at H/W=5, overall AP falls (33.3 → 31.8) while $AP_{vt}$ peaks at 17.9 (>17.8 baseline), consistent with the stated trade-off.
Efficiency: 36M params vs. 34M (Δ ≈ 5.88%); 358.7 GFLOPs − 327.5 GFLOPs = 31.2 GFLOPs (matches text).
DOI: 10.1145/3746027.3754861

Notes

This review is limited to text-extracted content; pixel-level image forensics (duplication, splicing, noise inconsistency) requires the original PDF or raw figures.
The strong cross-document consistency is a positive indicator, but absence of arithmetic inconsistency is necessary but not sufficient proof of integrity.
No recommendation for PubPeer post, author contact, or editorial notification based on current evidence.