Reproduce “A Reverse Causal Framework to Mitigate Spurious Correlations for Debiasing Scene Graph Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shuzhou Sun, Li Liu, Tianpeng Liu, Shuaifeng Zhi, Ming-Ming Cheng, Janne Heikkilä, Yongxiang Liu
A new framework for Scene Graph Generation (SGG) tackles prediction biases by re-engineering the training process with a reverse causal structure. It improves mean recall rates (mR@K) by up to 7.6% and better handles background relationships, maintaining overall performance.
Key insights
- Bias in Scene Graph Generation (SGG) stems not just from data distribution, but from spurious correlations inherent in the conventional training framework's causal structure.
- Reversing the causal chain in SGG training, making relationship features a 'collider', can break spurious correlations and effectively mitigate bias.
- Loss-driven adaptive sampling, inspired by active learning, can effectively rebalance training data by prioritizing categories with higher losses (often tail and foreground relationships).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.