Reproduce “ReMOVE: A Reference-free Metric for Object Erasure” 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

Aditya Chandrasekar, Goirik Chakrabarty, Jai Bardhan, Ramya Hebbalaguppe, Prathosh AP

ReMOVE is a new reference-free metric that evaluates the quality of object erasure in diffusion-based image editing by comparing the visual features of the inpainted area to its surrounding background. Experiments show ReMOVE correlates with reference-based metrics and aligns with human perception, addressing the challenge of assessing seamless object removal without ground truth images.

Key insights

  • Object erasure quality can be assessed by comparing visual features of the inpainted region to its surrounding background, rather than relying on a ground truth image.
  • Vision Transformers (ViT) pre-trained on segmentation tasks are effective for extracting features that capture both local textures and global spatial relationships crucial for evaluating seamless blending.
  • Distinguishing true object removal from object replacement is critical for an accurate erasure metric, a limitation addressed by focusing on feature similarity to the background.

Summary by alphaXiv.

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