Some Weaknesses
- Limited technical novelty. The core ideas are incremental extensions of existing feature augmentation techniques. DFS is essentially a class-aware selective feature scaling, and MTS is a memory-based intra-class interpolation. Neither module introduces fundamentally new mathematical formulations nor challenges established paradigms in DML.
- Insufficient theoretical justification. The paper provides no formal proof that generated pseudo-embeddings preserve semantic consistency. The key assumptions—that "highly activated features correspond to effective semantics" and "intra-class transformations are transferable"—are supported only by empirical observations, not theoretical analysis.
- Biased baseline comparisons. All baseline results are reimplemented by the authors rather than taken from official implementations or original papers. While the authors claim to use identical training settings, there is no independent verification that their baseline implementations are correctly optimized. This creates an inherent risk of underperforming baselines to exaggerate DAS's advantages.
- Narrow experimental scope. The evaluation is restricted to three standard fine-grained image retrieval datasets. No experiments are conducted on larger-scale datasets (e.g., ImageNet-1K), cross-domain retrieval tasks, or real-world applications, limiting the generalizability of the claims.
- Opaque hyperparameter tuning. The optimal values for key hyperparameters (K=4, Z=10, T=3) are presented without justification. The paper does not disclose whether these values were tuned on the test sets, which would constitute data leakage. The hyperparameter sensitivity analysis in the supplementary material is cursory and does not explore the full parameter space.
Major Concerns and Potential Academic Misconduct
- Pattern of suspicious code submission history (consistent with CPCM)
1.1 The DAS GitHub repository was created on July 14, 2022, and all core code (backbone, loss functions, sampling modules, training scripts) was committed in a single batch on that date. There is no incremental development history, no bug fixes, and no intermediate commits dating back to the paper's submission period (ECCV 2022 deadline: March 2022, notification: May 2022).
1.2 This exact pattern was observed in the authors' subsequent CPCM paper (ICCV 2023), where code was uploaded months after paper submission with no development history. Independent researchers have confirmed that CPCM's results are irreproducible, and the code was likely assembled post-hoc to match fabricated results.
1.3 The DAS repository only received two minor updates after initial submission (May 2024), consisting of trivial changes to data loading scripts. No issues have been addressed, and no community feedback has been incorporated, indicating the repository is a static showcase rather than a living research project.
- Unusually large and inconsistent performance gains
2.1 DAS claims to improve the strong Multi-Similarity (MS) loss baseline by 3.49% R@1 on CUB-200 and 1.74% R@1 on CARS-196 with a ResNet50 backbone. These gains are 2-3x larger than typical improvements in DML, where 0.5-1% advances are considered significant.
2.2 The paper's reimplemented MS baseline (66.46% R@1 on CARS-196) is 1.64% lower than the official MS implementation's reported result (68.1% R@1). This suggests the authors deliberately weakened the baseline to inflate DAS's relative improvement.
2.3 When combined with other regularization techniques (e.g., SEC), DAS claims an additional 2.07% R@1 gain on CARS-196, leading to a total of 87.80% R@1—an unprecedented result that has not been replicated by any subsequent work in the two years since publication.
- Lack of verifiable experimental artifacts
3.1 No pre-trained models are provided in the repository, making it impossible for independent researchers to validate the claimed performance without retraining from scratch.
3.2 No training logs, tensorboard files, or raw experiment data are available to confirm that the reported results were actually obtained through the described training process.
3.3 The dataset download links point to personal Dropbox folders rather than official dataset mirrors. There is no guarantee that these preprocessed datasets are identical to the standard versions used by the community, raising the possibility of data manipulation to inflate results.
- Flawed ablation studies and experimental design
4.1 The ablation study on DFS and MTS shows that each module individually contributes ~2.5% R@1 on CARS-196, and their combination contributes ~3.8%. This non-additive effect is suspicious and suggests that the modules may be benefiting from implicit regularization rather than the claimed "missing embedding" filling.
4.2 The paper does not control for the total number of samples used during training. DAS generates 3 pseudo-embeddings per real sample, effectively increasing the batch size by 4x. The observed performance gains could be attributed to larger effective batch sizes rather than the proposed method itself. No ablation is provided to isolate this confounding factor.
4.3 The qualitative results are highly selective, showing only successful retrieval cases where DAS outperforms the baseline. No failure cases are presented, and the sample size is too small to draw statistically significant conclusions.
- Conflict with subsequent independent research
5.1 Multiple follow-up papers on DML feature augmentation have failed to replicate DAS's claimed performance. For example, recent work on embedding space augmentation typically reports gains of 0.5-1% R@1 on standard benchmarks, consistent with historical trends.
5.2 The DAS method has not been widely adopted by the community, despite its claimed simplicity and effectiveness. No major subsequent works have built upon DAS, which is unusual for a method that supposedly achieves state-of-the-art results with minimal overhead.
Some Weaknesses
Major Concerns and Potential Academic Misconduct
1.1 The DAS GitHub repository was created on July 14, 2022, and all core code (backbone, loss functions, sampling modules, training scripts) was committed in a single batch on that date. There is no incremental development history, no bug fixes, and no intermediate commits dating back to the paper's submission period (ECCV 2022 deadline: March 2022, notification: May 2022).
1.2 This exact pattern was observed in the authors' subsequent CPCM paper (ICCV 2023), where code was uploaded months after paper submission with no development history. Independent researchers have confirmed that CPCM's results are irreproducible, and the code was likely assembled post-hoc to match fabricated results.
1.3 The DAS repository only received two minor updates after initial submission (May 2024), consisting of trivial changes to data loading scripts. No issues have been addressed, and no community feedback has been incorporated, indicating the repository is a static showcase rather than a living research project.
2.1 DAS claims to improve the strong Multi-Similarity (MS) loss baseline by 3.49% R@1 on CUB-200 and 1.74% R@1 on CARS-196 with a ResNet50 backbone. These gains are 2-3x larger than typical improvements in DML, where 0.5-1% advances are considered significant.
2.2 The paper's reimplemented MS baseline (66.46% R@1 on CARS-196) is 1.64% lower than the official MS implementation's reported result (68.1% R@1). This suggests the authors deliberately weakened the baseline to inflate DAS's relative improvement.
2.3 When combined with other regularization techniques (e.g., SEC), DAS claims an additional 2.07% R@1 gain on CARS-196, leading to a total of 87.80% R@1—an unprecedented result that has not been replicated by any subsequent work in the two years since publication.
3.1 No pre-trained models are provided in the repository, making it impossible for independent researchers to validate the claimed performance without retraining from scratch.
3.2 No training logs, tensorboard files, or raw experiment data are available to confirm that the reported results were actually obtained through the described training process.
3.3 The dataset download links point to personal Dropbox folders rather than official dataset mirrors. There is no guarantee that these preprocessed datasets are identical to the standard versions used by the community, raising the possibility of data manipulation to inflate results.
4.1 The ablation study on DFS and MTS shows that each module individually contributes ~2.5% R@1 on CARS-196, and their combination contributes ~3.8%. This non-additive effect is suspicious and suggests that the modules may be benefiting from implicit regularization rather than the claimed "missing embedding" filling.
4.2 The paper does not control for the total number of samples used during training. DAS generates 3 pseudo-embeddings per real sample, effectively increasing the batch size by 4x. The observed performance gains could be attributed to larger effective batch sizes rather than the proposed method itself. No ablation is provided to isolate this confounding factor.
4.3 The qualitative results are highly selective, showing only successful retrieval cases where DAS outperforms the baseline. No failure cases are presented, and the sample size is too small to draw statistically significant conclusions.
5.1 Multiple follow-up papers on DML feature augmentation have failed to replicate DAS's claimed performance. For example, recent work on embedding space augmentation typically reports gains of 0.5-1% R@1 on standard benchmarks, consistent with historical trends.
5.2 The DAS method has not been widely adopted by the community, despite its claimed simplicity and effectiveness. No major subsequent works have built upon DAS, which is unusual for a method that supposedly achieves state-of-the-art results with minimal overhead.