Communications of the ACM
In Plain Sight
Researchers are comparing the performance of human subjects with that of deep neural networks in visual searches.
Credit: The UC Santa Barbara Current
Researchers at the University of California, Santa Barbara (UCSB) have found deep neural networks can be superior to humans in conducting visual searches, because the networks do not overlook mis-scaled targets.
The networks search across whole scenes and use the visual characteristics of the object itself, while humans also employ the relationships between objects and their context within the scene to direct their eyes.
On the other hand, humans were found to be better than computer vision in verifying the presence of different target objects in real-world scenes that may or may not feature them, as networks returned a higher number of false positives.
The researchers note this mechanism is a useful approach of the human brain for processing scenes quickly, removing distractors, and lowering false positives.
"The findings might suggest ways to improve computer vision by implementing some of the tricks the brain utilizes to reduce false positives," says Middle East Technical University in Turkey professor Emre Akbas, who worked on the project while at UCSB.
From The UC Santa Barbara Current
View Full Article
Abstracts Copyright © 2017 Information Inc., Bethesda, Maryland, USA
No entries found