Are metrics measuring what they should? An evaluation of Image Captioning task metrics

Abstract

Image Captioning is a current research task to describe the image content using the objects and their relationships in the scene. Two important research areas converge to tackle this task: artificial vision and natural language processing. In Image Captioning, as in any computational intelligence task, the performance metrics are crucial for knowing how well (or bad) a method performs. In recent years, it has been observed that classical metrics based on $n$-grams are insufficient to capture the semantics and the critical meaning to describe the content in an image. Looking to measure how well or not the current and more recent metrics are doing, in this article, we present an evaluation of several kinds of Image Captioning metrics and a comparison between them using the well-known datasets, MS-COCO and Flickr8k. The metrics were selected from the most used in prior works; they are those based on $n$-grams, such as BLEU, SacreBLEU, METEOR, ROGUE-L, CIDEr, SPICE, and those based on embeddings, such as BERTScore and CLIPScore. We designed two scenarios for this: (1) a set of artificially built captions with several qualities and (2) a comparison of some state-of-the-art Image Captioning methods. Interesting findings were found trying to answer the questions: Are the current metrics helping to produce high-quality captions? How do actual metrics compare to each other? What are the metrics really measuring?