|
报告内容
框架
|
Recent years has witnessed the great
success of deep learning in a variety of vision tasks. In most cases, deep
learning is conducted in a supervised way. As for image search, since the
category number of potential objects is difficult to enumerate and the image
database is large, it is infeasible to collect sufficient annotated training
images as supervision for deep learning. As a result, most works on image
search simply leverage the activations from pre-trained deep learning model,
or just focus on some specific fine-grained tasks, such as landmark
retrieval. To this end, we explore deep learning in a pseudo-supervised
paradigm and orient it for image retrieval. We approach it from different
perspectives and propose three algorithms. Further, to automatically evaluate
the retrieval result quality, we propose a deep learning based quality
assessment method. Extensive experiments demonstrate the effectiveness and
potential of pseudo-supervised deep learning in
retrieval task.
|
