2000 character limit reached
Does SAM dream of EIG? Characterizing Interactive Segmenter Performance using Expected Information Gain (2404.16155v1)
Published 24 Apr 2024 in cs.CV, cs.IT, cs.LG, and math.IT
Abstract: We introduce an assessment procedure for interactive segmentation models. Based on concepts from Bayesian Experimental Design, the procedure measures a model's understanding of point prompts and their correspondence with the desired segmentation mask. We show that Oracle Dice index measurements are insensitive or even misleading in measuring this property. We demonstrate the use of the proposed procedure on three interactive segmentation models and subsets of two large image segmentation datasets.
Paper Prompts
Sign up for free to create and run prompts on this paper using GPT-5.
Top Community Prompts
Collections
Sign up for free to add this paper to one or more collections.
