Table 2.

Improvement in car/person detection obtained by augmenting state-of-the-art CNNs with predicted human-derived contextual expectations.

CNN	Target	CNN	CNN + Lklhd	CNN + yLocn	CNN + scale	CNN + Lklhd + yLocn + scale	CNN + all car & person ratings	Increase in %
RCNN3	C	82.4 ± 0.000	85 ± 0.000	82.5 ± 0.000	83.2 ± 0.000	85.5 ± 0.001	86.2 ± 0.0017	3.8
	P	80.6 ± 0.000	80.6 ± 0.001	81.5 ± 0.001	80.6 ± 0.000	80.4 ± 0.0018	82.0 ± 0.0023	1.4
Alexnet2	C	83.5 ± 0.000	85.8 ± 0.0012	83.5 ± 0.001	84.3 ± 0.0015	86.8 ± 0.0017	87.1 ± 0.0019	3.6
	P	73.4 ± 0.002	73 ± 0.0028	77.1 ± 0.0025	75.0 ± 0.0025	76.8 ± 0.0034	77.1 ± 0.0038	3.7

Each entry shows the cross-validated accuracy for detecting cars (C) or people (P) on novel scenes from the ADE20K dataset chosen from the same scene categories as in the human experiments (for details, see Supplementary Tables S2, S3 and S4). The best performing models are highlighted in bold. Columns indicate the kind of model used: the column marked CNN indicates the baseline accuracy of the deep neural network; the columns of the form “CNN + X” indicate accuracy for CNN augmented with feature X. Lklhd: predicted likelihood of target category object; xLocn: predicted horizontal location of target category object, yLocn: predicted vertical location of target category object; scale: overall bounding box area marked by subjects.