Pretrained model improves visual dialog performance.
problem Improving performance in visual dialog tasks.
method Pretrained ViLBERT model on vision-language datasets, fine-tuned on VisDial.
result Best model outperforms prior work by more than 1% on NDCG and MRR.