Reproduce “Grounding Image Composition with Label-Conditioned Visual Evidence” 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
Armin Dadras, Tingyu Lin, Robert Sablatnig, Franziska Proksa, Markus Seidl
KU-PCPE creates a shared human-annotated region target for all nine KU-PCP composition classes, asking which image regions support an applicable label. Label-conditioned SegFormer-B0 achieves .6636 mean Dice on ambiguity-aware expert-test consensus, compared with .8666 for a held-out expert. On the same 12-candidate bank, Qwen3-VL-32B scores .3946, while a target-aware selector reaches .6448, but is not a predictor.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.