. Author manuscript; available in PMC: 2021 Jun 22.

Published in final edited form as: Proc IEEE Comput Soc Conf Comput Vis Pattern Recognit. 2020 Aug 5;2020:190–199. doi: 10.1109/cvpr42600.2020.00027

Table 2: Comparing scanpath prediction algorithms.

(rows) using multiple scanpath metrics (columns) on the COCO-Search18 test dataset. In the case of Sequence Score and Multimatch, “Human” refers to an oracle method where one searcher’s scanpath is used to predict another searcher’s scanpath; “Human” for all other metrics refers to observed behavior.

	TFP-AUC ↑	Probability Mismatch ↓	Scanpath Ratio ↑	Sequence Score ↑	MultiMatch ↑
	TFP-AUC ↑	Probability Mismatch ↓	Scanpath Ratio ↑	Sequence Score ↑	shape	direction	length	position
Human	5.200	-	0.862	0.490	0.903	0.736	0.880	0.910
Random scanpath	0.795	4.407	-	0.295	0.869	0.558	0.849	0.849
Detector	4.046	1.166	0.687	0.414	0.877	0.676	0.853	0.863
Fixation heuristic	2.154	3.046	0.545	0.342	0.873	0.614	0.870	0.850
BC-CNN	3.893	1.328	0.706	0.409	0.880	0.669	0.865	0.874
BC-LSTM	1.702	3.497	0.406	0.324	0.834	0.567	0.818	0.770
IRL(Ours)	4.509	0.987	0.826	0.422	0.886	0.695	0.866	0.885