CLIPScore: A Reference-free Evaluation Metric for Image Captioning

Benchmark Model Rank Results
human-judgment-classification-on-pascal-50sRefCLIP-S#2Mean Accuracy: 83.1
human-judgment-classification-on-pascal-50sCLIP-S#3Mean Accuracy: 80.7
human-judgment-correlation-on-flickr8k-cfRefCLIP-S#2Kendall's Tau-b: 36.4
human-judgment-correlation-on-flickr8k-cfCLIP-S#3Kendall's Tau-b: 34.4
human-judgment-correlation-on-flickr8k-expertRefCLIP-S#3Kendall's Tau-c: 53.0
human-judgment-correlation-on-flickr8k-expertCLIP-S#4Kendall's Tau-c: 51.2