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

Table 2

Performance comparisons on MS COCO Karpathy test split under cross-entropy training.

Cross-entropy loss
Metric	BLEU1	BLEU2	BLEU3	BLEU4	METEOR	ROUGE-L	CIDEr	SPICE
NIC [16]	—	—	—	29.6	—	52.6	94.0	—
SCST [11]	—	—	—	30.0	25.9	53.4	99.4	—
Up-down [4]	77.2	—	—	36.2	27.0	56.4	113.5	20.3
RFNet [17]	76.4	60.4	46.6	35.8	27.4	56.8	112.5	20.5
HAN [20]	77.2	61.2	47.7	36.2	27.5	56.6	114.8	20.6
RAtt-Soft [29]	79.2	61.8	47.6	36.9	28.3	60.9	114.3	20.8
CFG	77.1	61.5	47.9	36.8	27.7	56.7	114.0	20.8

The best results (%) are highlighted in boldface. The symbol “—” indicates the results are not reported.