Call For Papers

Responsible Image Synthesis:

Fairness & Robustness: For example, collecting and preparing synthetic datasets that fairly represent different classes; simulating the most difficult and rare conditions to improve the robustness and generalizability of systems.
Bias & Ethics: For instance, implementing validation procedures to ensure that synthetic datasets do not contain unintentional bias; setting clear ethical guidelines to prevent the misuse of that data.
Privacy & Security: Examples include anonymization and de-identification to prevent data linking to real people; using synthetic data to enhance communication rounds in federated learning.
Reliability: Conducting experiments to test reliability and accuracy in various application contexts by comparing the results achieved using synthetic information with those obtained using real data; enabling a deeper understanding of the algorithms and decisions made in the generation process.

Learning from Synthetic Data:

Frameworks: Domain generalization by training models on a wide array of simulated conditions; addressing the domain adaptation challenge, and assessing the effectiveness of transfer learning techniques.
Strategies: Continual learning by generating dynamic datasets that reflect evolving conditions and novel challenges; one- or few-shot learning by generating diverse and comprehensive datasets from a limited number of real-world examples.
Theoretical Foundations: Development of benchmarks and validation protocols to evaluate the effectiveness of models trained on synthetic data; establishing standardized tests to ensure reliability and robustness across various synthetic datasets.

In the dynamic landscape of computer vision, synthetic data is emerging as a key resource to overcome data limitations and improve the accuracy and reliability of systems. Advanced generative models like diffusion models, GANs, and multimodal models can generate large amounts of data with different characteristics to address data insufficiency and bias, as well as privacy concerns, by removing sensitive information or generating data without it.
These advances are particularly significant in several areas, such as healthcare, where privacy and data protection laws limit access to real patient data. In addition, simulating rare medical disorders could improve model accuracy and generalizability. Furthermore, these data present a compelling opportunity for resource optimization by removing the need to store massive datasets; they can be generated and dynamically provided to learning models during training. This workshop aims to explore not only the favorable impacts of synthetic data, such as those just delineated, but also the challenges and risks associated with their potential misuse. Particularly in the area of security, synthetic data could be exploited to circumvent biometric recognition systems, undermining their effectiveness and enabling unauthorized access or fraudulent activities. Therefore, there is an imperative not only to generate increasingly high-quality data but also to develop robust algorithms for their detection.

The workshop welcomes contributions on all topics related to synthetic data in the field of computer vision, focused (but not limited to):

Generative Models:

Stable Diffusion Models: Such as personalization, conditional generation, guidance, and controllability; innovative approaches for training or architectures.
3D Models: Overcoming challenges related to shape diversity, structure, and object complexity; exploring how they can be integrated into VR and AR applications.
Deepfakes: Developing algorithms and systems to identify and neutralize deepfake content to prevent misinformation and protect individual identities; forensic techniques for analysis and attribution of deepfake videos.

Applications:

Medical Image Synthesis:
- Generation of Synthetic Diagnostic Images: Improving the realism, diversity, and clinical relevance of synthetic medical images to aid in training and evaluating diagnostic systems; ethical implications of using synthetic data in healthcare: patient privacy concerns, and strategies to mitigate potential risks.
- Simulation of Pathological Variants for Medical Model Training: Incorporating expert knowledge into a machine learning model to define a fine-tuning objective; techniques for tailoring synthetic images conditioned on clinical concepts.
- Synthetic Data for Disease Progression Modeling: Frameworks that capture the temporal evolution of diseases; simulations of the impact of potential treatments and interventions, providing a controlled environment for testing novel medical strategies.
Synthetic Biometrics:
- Innovative Synthesis of Biometric Data: Development of advanced techniques for synthetic biometric generation; identification of core technical challenges.
- Label Generation: Implementation of automatic annotation techniques; ensuring label accuracy and consistency.
- Quality Assessment: Methods to evaluate synthetic biometric data quality and realism; comparison tools and metrics against real data.
Others.

Images generated using DALL·E