Geng report: 'Road Obstacle Detection Fusing Hybrid Attention and Detection Head' (Li Yujuan et al.) — academic-integrity review

Academic fraud report · Geng Detector

Summary

This review examines allegations of academic misconduct in the article 'Road Obstacle Detection Fusing Hybrid Attention and Detection Head' by Li Yujuan and colleagues, published in Science Technology and Engineering (DOI: 10.12404/j.issn.1671-1815.2500190). The verdict is 'substantiated (实锤)' based on two primary findings. First, Table 1 (ablation study) reports parameter counts of 2.81M for adding AKConv alone, 2.89M for adding D-DySample alone, but 10.43M when both modules are added together — a non-additive jump that is mathematically inconsistent with module-stacking principles, suggesting either undisclosed backbone changes or fabricated figures. Second, the proposed ADMH-YOLOv8 model uses 9.65M parameters and 20.5G FLOPs yet is benchmarked against lightweight nano/tiny models (YOLOv5n, YOLOv9-t, YOLOv10, YOLOv11), making the mAP comparison misleading ('horse-racing' style benchmarking). Secondary issues include duplicated text in Section 3.2, sub-0.5% mAP improvements reported without statistical measures, and reliance on an expired 2017 NSFC grant. Limitations: numeric values are reproduced as cited; no raw image forensics was performed.

Verdict

🔴 Substantiated — Major unresolved mathematical inconsistency in Table 1 plus misleading benchmark-scale selection. Both issues are independently sufficient to warrant editorial investigation.

Key findings

Impossible parameter-count jump in ablation (Finding 1). Table 1 shows that adding AKConv alone raises parameters to 2.81M and D-DySample alone to 2.89M, but adding both together produces 10.43M — roughly 3.6× the baseline (2.87M), violating additive modular logic.
Unfair ('horse-racing') comparison (Finding 2). The proposed ADMH-YOLOv8 uses 9.65M parameters and 20.5G FLOPs (Table 3), yet is benchmarked against YOLOv5n (1.68M), YOLOv9-t (2.50M), YOLOv10 (2.57M), and YOLOv11 (2.46M). Higher mAP from a model 3–4× larger is expected and does not validate the claimed modules.
Duplicated text in Section 3.2 (Finding 3). The phrase '类别信息则标明了障碍物的类别' appears twice consecutively, suggesting copy-paste drafting errors.
Statistical weakness in ablation gains (Finding 4). Reported mAP50 improvements between configurations are only 0.1%–0.4% (e.g., 91.8% → 91.9% → 92.1% → 92.2%) with no mean ± SD across runs, falling within typical training-run variance.
Aging funding support (Finding 5). Only listed grant is NSFC 51765007, normally a 2017 project cycle; paper submitted 2025-01-08.

Evidence highlights

Table 1 parameter counts: baseline 2.87M; +AKConv 2.81M; +D-DySample 2.89M; +both 10.43M (non-additive, ~3.6×).
Table 3 reported metrics: ADMH-YOLOv8 — 9.65M params, 20.5G FLOPs; YOLOv5n 1.68M; YOLOv9-t 2.50M; YOLOv10 2.57M; YOLOv11 2.46M.
mAP50 ablation range: 91.8%–92.2% (Δ ≤ 0.4%), no SD reported.
Duplicated phrase in §3.2: '类别信息则标明了障碍物的类别.'
Submission date 2025-01-08; NSFC grant 51765007 (2017 cycle).
DOI: 10.12404/j.issn.1671-1815.2500190.

Notes

All numeric values are reproduced as cited in the source report; the original PDF was not independently re-extracted.
No image-level (Western blot-style) forensics was performed; pixel reuse or splicing cannot be ruled in or out.
Timeline reference to YOLOv11 (~Oct 2024) is consistent with the Jan 2025 submission.
The report explicitly recommends contacting the authors for training logs/configs, posting on PubPeer, and notifying the journal editorial office.
Final determination of misconduct requires formal institutional investigation.