Reproduce “Object Discovery with a Copy-Pasting GAN” 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

Relja Arandjelović, Andrew Zisserman

A Copy-Pasting GAN (CP-GAN) learns unsupervised object segmentation by training a generator to produce realistic composite images through copying objects, while a discriminator aims to identify these fakes. This discriminative approach addresses limitations of generative models by achieving high object discovery performance on cluttered and varied datasets, including 100% on "NoisySquares" and 40.3% on "Flying Chairs".

Key insights

  • Object segmentation emerges as an optimal strategy for a generator attempting to create realistic image composites by copying objects from one image and pasting them into another.
  • Unsupervised object discovery can be effectively framed as a discriminative adversarial game, offering an alternative to traditional generative scene reconstruction methods.
  • A suite of anti-shortcut mechanisms, discriminator blurring, auxiliary mask prediction, and grounded fakes are crucial for stabilizing GAN training and preventing trivial solutions in this complex task.

Summary by alphaXiv.

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