Transient Attributes for High-Level Understanding and Editing of Outdoor Scenes

摘要

We live in a dynamic visual world where the appearance of scenesrnchanges dramatically from hour to hour or season to season. Inrnthis work we study “transient scene attributes” – high level propertiesrnwhich affect scene appearance, such as “snow”, “autumn”,rn“dusk”, “fog”. We define 40 transient attributes and use crowdsourcingrnto annotate thousands of images from 101 webcams. Wernuse this “transient attribute database” to train regressors that canrnpredict the presence of attributes in novel images. We demonstraterna photo organization method based on predicted attributes. Finallyrnwe propose a high-level image editing method which allows a userrnto adjust the attributes of a scene, e.g. change a scene to be “snowy”rnor “sunset”. To support attribute manipulation we introduce a novelrnappearance transfer technique which is simple and fast yet competitivernwith the state-of-the-art. We show that we can convincinglyrnmodify many transient attributes in outdoor scenes.