Core Concepts
The author explores the importance of considering both local and global attribute variations in evaluating segmentation models, highlighting the impact on performance.
Abstract
The content discusses a pipeline for editing visual attributes of real images while preserving original segmentation labels. A benchmark is constructed to evaluate segmentation models' robustness to different attribute variations. Results show vulnerability to object attribute changes and the importance of considering local attributes for improved robustness. The quality of edited images is assessed through comparisons with existing benchmarks and image editing methods.
Stats
Material: wood, stone, metal, paper
Color: violet, pink
Pattern: dotted, striped
Style: snowy, painting, sketch
mIoU drop ↓: 15.33%, 22.06%, 31.19%, 21.45%, 21.82%, 26.32%, 34.99%, 34.45%, 28.18%
Quotes
"We argue that local attributes have the same importance as global attributes."
"Performance declines most on object material variations."