SIBGRAPI Paper Reviews¶

Review 1¶

Lack of novelty
"The 4.4% improvement in accuracy is true but does not tell the majority case story."
"The aggregate score, currently a simple mean, could be elevated to a more advanced level by employing sophisticated aggregator functions, inspiring research in this area."

The authors fail to position their work in relation to the existing literature.
Choice of related work should have been based on gender classification on FairFace.
The paper fails to convey to the reader the current state of the art.
Methodology imprecise in describing dataset evaluation protocol (entire dataset or test set?)
Introduction fails to define the research problem.
FairFace dataset is absent from abstract and introduction.
Conclusion overly lengthy
Explore whether FairFace was used to train WIT or LAION.

"Gadre and colleagues [14] introduced CommonPool" should be "Gadre et al".
Introduction should formalize the problem simillarly to section III.A
Research questions should be stated more explicitly.
Incoherent "For all models, the [...] achieved the highest accuracy for some models"

Broaden the scope of related works to include studies that specifically address gender classification using the FairFace dataset and CLIP-based models.
Provide more detailed explanations about the use of the FairFace dataset, particularly whether the entire dataset or only the test partition was used.
Introduction should clearly formalize the problem we are addressing and explicitly state the research questions and objectives.
Clarify overlaps between training datasets and FairFace dataset, to address potential data leakage concerns.
Condense conclusion to focus on key findings, implications and future research directions.
Add tables or figures that compare our results with those from previous studies on similar tasks.
Provide public repo.