Context-Fused Guidance for Image Captioning Using Sequence-Level Training

Table 1

Statistics of the MS COCO dataset.

Split	Default		Karpathy
Subset	Image	Caption	Image	Caption
Training	82,783	414,113	113,287	566,738
Validation	40,504	202,654	5000	25,010
Test	40,775	—	5000	25,010

The symbol “—” indicates the data are not public.