Reproduce “Fine-Grained Scene Graph Generation via Sample-Level Bias Prediction” on OpenResearch

About this paper

Yansheng Li, Tingzhu Wang, Kang Wu, Linlin Wang, Xin Guo, Wenbin Wang

Wuhan University researchers developed a Sample-Level Bias Prediction method leveraging a Bias-Oriented Generative Adversarial Network (BGAN) to address the long-tailed distribution in Scene Graph Generation (SGG). Their approach produces fine-grained relationship predictions by dynamically correcting biased logits, leading to superior Average@K scores and increased recall for rare relationship categories across various SGG benchmarks.

Key insights

• The union region of an object pair contains unique contextual visual information that can be leveraged to predict a precise, sample-specific bias for relationship correction.
• Learning these sample-specific biases allows for significant refinement of initial coarse-grained relationship predictions towards more accurate fine-grained ones.
• A Bias-Oriented Generative Adversarial Network (BGAN) can effectively learn complex, non-linear biases from visual context and global priors for improved debiasing.

Summary by alphaXiv.