DLC-Bench
收藏资源简介:
DLC-Bench是由NVIDIA提出的,用于评估详细局部化图像和视频字幕生成的基准。该数据集通过半监督学习数据管道DLC-SDP生成,它利用高质量的分段注释和未标记的网络图像来丰富区域描述。DLC-Bench的设计目的是为了评估详细局部化字幕,而无需依赖参考字幕,它为模型提供了多种粒度的评估,包括关键词级别、短语级别和详细的 multisentence 局部化图像和视频字幕生成。
DLC-Bench, proposed by NVIDIA, is a benchmark for evaluating detailed grounded image and video captioning. This benchmark is generated via the semi-supervised learning data pipeline DLC-SDP, which leverages high-quality segmentation annotations and unlabeled web images to enrich regional descriptions. DLC-Bench is designed to evaluate detailed grounded captioning without relying on reference captions, and it provides models with evaluations across multiple granularities, including keyword-level, phrase-level, and detailed multi-sentence grounded image and video captioning.




