Evaluating the WordsEye Text-to-Scene System: Imaginative and Realistic Sentences

We describe our evaluation of the WordsEye text-to-scene generation system. We address the problem of evaluating the output of such a system vs. simple search methods to find a picture to illustrate a sentence. To do this, we constructed two sets of test sentences: a set of crowdsourced imaginative sentences and a set of realistic sentences extracted from the PASCAL image caption corpus (Rashtchian et al., 2010). For each sentence, we compared sample pictures found using Google Image Search to those produced by WordsEye. We then crowdsourced judgments as to which picture best illustrated each sentence. For imaginative sentences, pictures produced by WordsEye were preferred, but for realistic sentences, Google Image Search results were preferred. We also used crowdsourcing to obtain a rating for how well each picture illustrated the sentence, from 1 (completely correct) to 5 (completely incorrect). WordsEye pictures had an average rating of 2.58 on imaginative sentences and 2.54 on realistic sentences; Google images had an average rating of 3.82 on imaginative sentences and 1.87 on realistic sentences. We also discuss the sources of errors in the WordsEye system.