Teaching robots to see
Source: Victoria University
Wellington Cable Car image with different salient object detection algorithms
Syed Saud Naqvi, a PhD student from Pakistan, is working on an algorithm to help computer programmes and robots to view static images in a way that is closer to how humans see.
Quick assumption: Head-mounted cameras reveal actions without showing the identity of the filmmaker. Message from computer vision experts: Not so fast.
Yedid Hoshen and Shmuel Peleg of the Hebrew University of Jerusalem are authors of the study, "Egocentric Video Biometrics," submitted last month on arXiv. They said that egocentric video is different from hand-held video: The camera is on the user, attached to the user's head, and video can be recorded at any time such as when the user is walking. It is assumed that the anonymity of the wearer can be preserved even when the video is publicly distributed, that the camera does not record images of the user, whose identity is hidden. After reading this paper, one may not be so sure. The authors said that it is possible for user identity to be recovered. They detailed their method for learning biometrics from head-worn cameras. "We show that motion features in egocentric video provide biometric information, and the identity of the user can be determined quite reliably from a few seconds of video," they wrote. They stated that It has been found that people can be distinctly identified by biometric characteristics such as height, stride length and walking speed.
They extracted biometric information from egocentric video, concentrating on video recorded when the user is walking. Our method relied on biometric information implicit in camera motion, which can identify users accurately. "Egocentric video suffers from bouncy and unsteady motion caused by user head and body motion," they wrote. "Although usually a nuisance, we show that this information can be useful for biometric feature extraction and consequently for identifying the user."
They talked about "Convolutional Neural Network (CNN) architectures on coarse optical flow" as the way in which biometrics may be extracted. They learned biometric features and classifiers using the CNN architecture, which includes layers corresponding to biometric feature extraction and classification. Biometric features were learned automatically using CNNs―these architectures were shown to generalize and improve on physically motivated hand designed features.
MIT Technology Review, in the "Emerging Technology From the arXiv" explained their method further, saying they focused on the optical flow in the videos―the pattern of motion of objects, edges and surfaces in the video from frame to frame. "This can be extracted relatively quickly from sequences just a few seconds long. They used 80 percent of the data extracted in this way to train a neural network to spot the unique pattern of optical flow associated with each user. They then used the remaining 20 percent of the data to test how accurately the trained network could spot each individual filmmaker, while comparing the technique with other machine learning approaches."
Two lessons from their study are, first, that this sort of biometric identification can be used to prevent unauthorized use of these cameras if the camera is programmed to work only by the one user. The second lesson is that this biometric method could identify a filmmaker. This biometric approach could be used to identify the filmmaker who had otherwise hoped to keep identity unknown. "An important message in this paper," they said, "is that people should be aware that sharing egocentric video will compromise their anonymity."
This research was supported by Intel ICRI-CI, by Israel Ministry of Science, and by the Israel Science Foundation.
| }
|