Reproduce “Panoptic Scene Graph Generation with Semantics-Prototype Learning” 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
Li Li, Wei Ji, Yiming Wu, Mengze Li, You Qin, Lina Wei, Roger Zimmermann
The paper introduces ADTrans (Adaptive Data Transfer), a framework addressing biased predicate annotations in Panoptic Scene Graph Generation (PSG) datasets. It adaptively transfers biased labels and incorporates missed positive samples by learning robust, unbiased predicate representations through specialized contrastive training and dynamic prototype learning, leading to improved performance of PSG models.
Key insights
- Human-annotated datasets for Panoptic Scene Graph Generation (PSG) often contain biased predicate annotations due to varying language preferences and semantic overlaps, hindering model learning.
- Learning robust, unbiased predicate representations by aligning textual and visual features and minimizing variance (IRM) is crucial for identifying and correcting biased data.
- Intelligently transferring biased annotations and incorporating missed positive samples can create a more consistent and informative dataset, outperforming simple data removal or prior transfer methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.