Localizing Objects While Learning Their Appearance

摘要: Learning a new object class from cluttered training images is very challenging when the location of instances unknown. Previous works generally require objects covering large portion images. We present novel approach that can cope with extensive clutter as well scale and appearance variations between instances. To make this possible we propose conditional random field starts generic knowledge then progressively adapts to class. Our simultaneously localizes while learning an model specific for demonstrate on PASCAL VOC 2007 dataset. Furthermore, our method enables train any state-of-the-art detector in weakly supervised fashion, although it would normally annotations.