Основні поняття
Enhancing dataset quality and model performance through the renovation of names in open-vocabulary segmentation benchmarks.
Анотація
The paper introduces RENOVATE, a framework to improve name quality in open-vocabulary segmentation benchmarks. By generating more precise names aligned with visual segments, datasets are upgraded and models trained with these renovated names show significant performance improvements. Human studies confirm the superiority of renovated names over original ones. The method involves candidate name generation using context information, training a renaming model, and automating name quality evaluation using pre-trained models.
Статистика
Names lead to up to 16% relative improvement from original names on various benchmarks.
Renovated names enable training of stronger open-vocabulary segmentation models.
Upgraded benchmarks contain 4 to 5 times more classes than original ones.
Models trained with renovated names show improvements in PQ, AP, and mIoU metrics.
FC-CLIP model with RENOVATE names achieves state-of-the-art performances on ADE20K and Cityscapes.
Цитати
"Names are essential to both human cognition and vision-language models."
"In this work, we focus on renovating the names for open-vocabulary segmentation benchmarks."
"Our results demonstrate that our renovated names are preferred in 82% cases."
"Our upgraded benchmarks allow us to conduct more fine-grained analysis on open-vocabulary model abilities and biases."
"Our results show that RENOVATE class names are of high quality since they help boost the performance of open-vocabulary models."