CLEVR-Humans

We collect a new dataset of human-posed free-form natural language questions about CLEVR images. Many of these questions have out-of-vocabulary words and require reasoning skills that are absent from our model’s repertoire