Reproduce “Say As You Wish: Fine-grained Control of Image Caption Generation with Abstract Scene Graphs” 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

Shizhe Chen, Qin Jin, Peng Wang, Qi Wu

Researchers developed a method to generate image captions with fine-grained control over content and detail using Abstract Scene Graphs (ASGs). The ASG2Caption model interprets these label-agnostic graphs, producing captions that closely adhere to specified attributes and relationships, and significantly increases caption diversity compared to previous models.

Key insights

  • Fine-grained control over caption content and detail can be achieved using abstract, label-agnostic graph structures, simplifying user input and avoiding the need for complex semantic parsing.
  • A 'role-aware' encoding of graph nodes, combined with multi-relational graph convolutions, effectively interprets user intentions and visual features within the context of the abstract graph.
  • Dynamically fusing graph content and structural flow attention, alongside a graph updating mechanism, enables the decoder to comprehensively and non-redundantly describe complex graph information while adhering to the specified structure.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.