Seeing is Worse than Believing: Reading\ud People’s Minds Better than Computer-Vision\ud Methods Recognize Actions

摘要

We had human subjects perform a one-out-of-six class action\udrecognition task from video stimuli while undergoing functional magnetic\udresonance imaging (fMRI). Support-vector machines (SVMs) were\udtrained on the recovered brain scans to classify actions observed during\udimaging, yielding average classification accuracy of 69.73% when tested\udon scans from the same subject and of 34.80% when tested on scans\udfrom different subjects. An apples-to-apples comparison was performed\udwith all publicly available software that implements state-of-the-art action\udrecognition on the same video corpus with the same cross-validation\udregimen and same partitioning into training and test sets, yielding classification\udaccuracies between 31.25% and 52.34%. This indicates that\udone can read people’s minds better than state-of-the-art computer-vision\udmethods can perform action recognition.