RcSGG samples relationship features based on training losses and object-pair diversity to reduce head-tail and foreground-background biases in scene graph generation, achieving state-of-the-art mean recall on four benchmarks.
Visual commonsense r-cnn,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
A Reverse Causal Framework to Mitigate Spurious Correlations for Debiasing Scene Graph Generation
RcSGG samples relationship features based on training losses and object-pair diversity to reduce head-tail and foreground-background biases in scene graph generation, achieving state-of-the-art mean recall on four benchmarks.