Unbiased SGG

Unbiased scene graph generation (SGG) aims to create accurate and fair representations of scenes by mitigating biases in existing datasets and algorithms. Current research focuses on developing new evaluation metrics that address inherent imbalances in predicate categories and improving model robustness through techniques like noisy label correction and weighted loss functions within various model architectures, including transformer-based and gradient boosting decision tree approaches. Addressing these biases is crucial for improving the fairness and reliability of SGG models, leading to more accurate and equitable applications in areas such as image understanding and healthcare.