核心概念
Enhancing open-vocabulary segmentation models through precise name renovation.
摘要
The study focuses on the importance of names in open-vocabulary segmentation benchmarks. It introduces the RENOVATE framework to improve dataset quality by renovating names. The process involves generating candidate names, training a renaming model, and validating the quality of renovated names through human preference studies. Results show significant improvements in benchmark challenges and model performance with renovated names.
Introduction:
- Names play a crucial role in open-vocabulary segmentation benchmarks.
- The RENOVATE framework aims to enhance dataset quality by improving naming precision.
- Human cognition relies heavily on categorization facilitated by descriptive labels.
Methodology:
- Candidate name generation involves leveraging context information for better results.
- The renaming model uses text queries for improved segment-name matching.
- Validation includes human preference studies and rigorous verification processes.
Experiments:
- Upgraded benchmarks demonstrate increased challenge levels and realism with more classes.
- Training models with renovated names leads to improved segmentation performance.
- Automated evaluation using pre-trained models confirms the effectiveness of RENOVATE names.
统计
人間の認知において、名前は重要であり、RENOVATEフレームワークを導入してデータセットの品質を向上させることが目的です。
引用
"Names are essential to both human cognition and vision-language models."
"Our renovated names lead to up to 16% relative improvement from the original names."