Review Kinship Verification SOTA

[vitalwarley]

• [ ] #57
• [ ] #58

[vitalwarley]

Solving the Families In the Wild Kinship Verification Challenge by Program Synthesis

Encontrei no IEEE e inspecionei por curiosidade.

vuvko é o que eu estava tentando reproduzir (#24). Há outro melhor: zxm123. ~Quem será?~ É o #26.

[vitalwarley]

De The 5th Recognizing Families in the Wild Data Challenge: Predicting Kinship from Faces, temos

onde

Podemos ver que TeamCNU, à frente do vuvko, é o melhor. Seu resultado é o mesmo que citei no comentário anterior. Abaixo colo a referência

X. Zhang, M. Xu, X. Zhou, and G. Guo. Supervised contrastive learning for facial kinship recognition. In Conference on Automatic Face and Gesture Recognition (FG), 2021.

[vitalwarley]

Supervised Contrastive Learning for Facial Kinship Recognition

Não consegui o PDF, mas encontrei o repositório. Implementado em PyTorch!

[vitalwarley]

A survey on kinship verification (Wang et al, 2023)

[vitalwarley]

A survey on kinship verification (Wang et al, 2023)

in this paper, we propose a multi-modal dataset for kinship verification containing a wider range of age variations than existing datasets. The newly collected Nemo-kinship dataset con- sists of 4216 videos of 85 families with 248 individuals.

@tfvieira, parece bastante promissor para responder nossa RQ.

[vitalwarley]

A survey on kinship verification (Wang et al, 2023)

Deep learning-based methods show good performance in solving extrinsic challenges. One of the extrinsic challenges lies in that ”Kinship verification databases are born with unbalanced data” [117]. A kinship dataset of N pairs of positive samples contains N(N - 1) potential negative pairs leading to a large unbalance. However, most of the current methods only use N negative pairs.

No caso do SOTA, temos, na verdade, 2 * (batch_size^2 - batch_size) pares negativos. Mais detalhes aqui.

[vitalwarley]

A survey on kinship verification (Wang et al, 2023)

[vitalwarley]

A survey on kinship verification (Wang et al, 2023)

in this paper, we propose a multi-modal dataset for kinship verification containing a wider range of age variations than existing datasets. The newly collected Nemo-kinship dataset con- sists of 4216 videos of 85 families with 248 individuals.

@tfvieira, parece bastante promissor para responder nossa RQ.

[vitalwarley]

Criei a #57 e #58 para nos ajudar a decidir sobre o tema de uma revisão sistemática. É requisito para disciplina PPGI021 e também foi sugerido pelo Bruno Georgevich na call do grupo de pesquisa ZOIOZ na última sexta (02/02).

[vitalwarley]

Via ChatGPT, fiz um brainstorming de perguntas de pesquisa para usarmos. A nossa atual pergunta reflete as perguntas nos itens 1 e 3.

---

In the domain of kinship recognition within computer vision, there are numerous intriguing research questions that can drive forward the understanding and technological advancements in this field. Here are a few research questions that could form the basis of impactful studies:

1. Algorithmic Performance Across Different Kin Relationships:

• How do current kinship recognition algorithms perform in distinguishing between different types of kin relationships (e.g., parent-child vs. sibling relationships) across diverse datasets?
• What is the impact of age differences and gender combinations on the accuracy of kinship recognition algorithms?

2. Cross-Cultural and Cross-Racial Efficacy:

• How effective are kinship recognition systems across different ethnic and cultural groups, and what are the challenges in developing culturally agnostic models?
• Can kinship recognition systems maintain high accuracy levels across diverse racial backgrounds, considering the variations in facial features and skin tones?

3. Temporal Dynamics in Kinship Recognition:

• How do facial aging processes affect the performance of kinship recognition algorithms over time, particularly for long-term kinship verification?
• What are the most effective methods for modeling and compensating for age progression in kinship recognition tasks?

4. Influence of Environmental and Contextual Factors:

• To what extent do environmental factors (e.g., lighting, background, image quality) and contextual information (e.g., clothing, hairstyle) impact the accuracy of kinship recognition systems?
• How can kinship recognition systems be optimized to perform under unconstrained environmental conditions?

5. Advanced Feature Extraction Techniques:

• What novel feature extraction techniques can enhance the performance of kinship recognition models, particularly in distinguishing subtle familial traits?
• How can deep learning architectures be leveraged or modified to capture complex genetic and phenotypic patterns indicative of kinship?

6. Integration of Multimodal Data:

• Can the integration of multimodal data sources (e.g., facial images, voice data, behavioral traits) improve the accuracy and robustness of kinship recognition systems?
• What are the challenges and ethical considerations in using multimodal data for kinship verification?

7. Ethical and Privacy Implications:

• What are the ethical implications of deploying kinship recognition technologies, particularly in terms of privacy, consent, and potential misuse?
• How can kinship recognition systems be designed to adhere to ethical AI principles and ensure the protection of individuals' privacy?

8. Comparison with Human Perception:

• How does the accuracy of automated kinship recognition systems compare with human abilities to recognize kinship relationships based on facial features?
• What insights from human perceptual studies of kinship can be applied to improve algorithmic approaches?

These research questions not only aim to push the technological boundaries of kinship recognition but also address the critical ethical, privacy, and societal concerns associated with this technology. Addressing these questions could lead to significant advancements in the field, contributing to more accurate, fair, and responsible kinship recognition systems.