A Reverse Causal Framework to Mitigate Spurious Correlations for Debiasing Scene Graph Generation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing two-stage Scene Graph Generation (SGG) frameworks use a causal chain structure.
• This structure can lead to spurious correlations and biases, such as tail relationships being predicted as head ones.

Purpose of the Study:

• To propose a new framework, RcSGG, that reconstructs the causal structure to mitigate biases in SGG.
• To improve the accuracy and reliability of scene graph generation.

Main Methods:

• Reconstructing the causal chain into a reverse causal structure with classifier inputs as the confounder.
• Employing Active Reverse Estimation (ARE) to estimate reverse causality.
• Utilizing Maximum Information Sampling (MIS) to enhance reverse causality estimation.

Main Results:

• The proposed RcSGG framework theoretically mitigates spurious correlations and induced biases.
• Comprehensive experiments demonstrate state-of-the-art mean recall rates on popular benchmarks.
• The framework shows effectiveness across diverse SGG approaches.

Conclusions:

• RcSGG offers a robust solution to inherent biases in traditional SGG frameworks.
• The reverse causal structure effectively addresses limitations of the standard causal chain.
• This work advances the field of scene graph generation with improved performance.