2019-02-01 Esemény a Facebookon
Jordán Sándor előadásában
Scene understanding with Deep Learning
Jelenet értelmezés Deep Learning megoldással
Időpont : 2019. 02. 01. 16:00
Az Óbudai Egyetem Bécsi úti kampuszán, a Neumann János Informatikai Kar épülete, F09-es előadó.
Mindenkit nagy szeretettel várunk!
Decades of research have shown that it is possible to extract the gist of a scene very easily. Other research has shown that some categories of objects,
people and animals in particular, can also be detected from very brief exposures. It may even be possible to infer from the features of a scene some properties of these objects, such as a person’s gender and perhaps their emotion.
Actions can also be characterized by relatively simple features.
Combining these pieces of information might therefore have led one to a simple ‘gist’ level of interpretation of a scene.
Scene understanding doesn’t stop at the gist level.
When an out of order event happens with a person, the pose becomes unusual, deviating from poses that we have associated with typical human activities, like sitting.
Moreover, some particular events might result in the loss of distinctive features from a view, like a wing of an airplane shaded by a building. These factors would make it harder to recognize a pattern of pixels as an object, which would obviously affect whether these details would be included in a scene interpretation following a brief exposure.
Avoidance is also a far more difficult action to detect than sitting, characterized by a relatively subtle shift in body.
A scene is telling us a story. Some stories are simpler than others, and for more ones it is not straightforward to follow the various characters and they role in the high level picture. The same is true for scenes. Every story must begin with a context. The characters of a story must be planted in some conceptual, and sometimes just perceptual, background. This is the gist of a scene. On the top of the gist come the characters, that are the objects of a scene. The next level is the specification of relationships between these objects, both with respect to each other and the background, defining the actions of a scene. Some actions can only be understood in the collective. The set of such actions build up the event of the scene. Finally, a good story should draw a person (or an artificially intelligent agent) into the scene to contemplate on it and think over it at deeper levels, which ultimately represent the understanding of a scene.
The talk will guide through the state of the art in visual object, gist, action, event and scene detection.