MIT Press

Table 1:

CiC-Language Generalization. NES on real-world language from the Chairs-in-Context (CiC) dataset. *SG architectures from Achlioptas et al. (2019) are the previously reported state-of-the-art method. NES⁺ grounds sub-events on the feature grid input. -SN indicates ShapeNet pre-trained features.

Method	Input	Listener Acc.
Majority	N/A	0.333
*SG-NoAttn	VGG16-SN	0.812 ± 0.008
*SG-Attn	VGG16-SN	0.817 ± 0.008
LSTM-Attn	VGG16-SN	0.731 ± 0.012
PoE	VGG16-SN	0.752 ± 0.009
NMN	VGG16-SN	0.763 ± 0.023
MAC	VGG16-SN	0.818 ± 0.013
NES	VGG16	0.842 ± 0.005
NES	VGG16-SN	0.856 ± 0.005
NES	Res101	0.853 ± 0.011
NES⁺	Res101	0.870 ± 0.009

Method	Input	Listener Acc.
Majority	N/A	0.333
*SG-NoAttn	VGG16-SN	0.812 ± 0.008
*SG-Attn	VGG16-SN	0.817 ± 0.008
LSTM-Attn	VGG16-SN	0.731 ± 0.012
PoE	VGG16-SN	0.752 ± 0.009
NMN	VGG16-SN	0.763 ± 0.023
MAC	VGG16-SN	0.818 ± 0.013
NES	VGG16	0.842 ± 0.005
NES	VGG16-SN	0.856 ± 0.005
NES	Res101	0.853 ± 0.011
NES⁺	Res101	0.870 ± 0.009