- 1、本文档共23页,可阅读全部内容。
- 2、原创力文档(book118)网站文档一经付费(服务费),不意味着购买了该文档的版权,仅供个人/单位学习、研究之用,不得用于商业用途,未经授权,严禁复制、发行、汇编、翻译或者网络传播等,侵权必究。
- 3、本站所有内容均由合作方或网友上传,本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺!文档内容仅供研究参考,付费前请自行鉴别。如您付费,意味着您自己接受本站规则且自行承担风险,本站不退款、不进行额外附加服务;查看《如何避免下载的几个坑》。如果您已付费下载过本站文档,您可以点击 这里二次下载。
- 4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等,请点击“版权申诉”(推荐),也可以打举报电话:400-050-0827(电话支持时间:9:00-18:30)。
查看更多
Hierarchical Novelty Detection for Visual Object Recognition
8
1 Abstract
0 Deep neural networks have achieved impressive success Test image:
2
in large-scale visual object recognition tasks with a prede-
n fined set of classes. However, recognizing objects of novel True label: Siamese cat Angora cat Dachshund Pika
Prior works: Siamese cat novel novel novel
u Ours: Siamese cat novel cat novel dog novel animal
J classes unseen during training still remains challenging.
The problem of detecting such novel classes has been ad- animal
5
1 dressed in the literature, but most prior works have focused cat dog
on providing simple binary or regressive decisions, e.g., the
] output would be “known,” “novel,” or corresponding con- Persian cat Siamese cat Pomeranian Welsh corgi
V fidence intervals. In this paper, we study more informative
C novelty detection schemes based on a hierarchical classifi-
s. cation framework. For an object of a novel class, we aim
c for finding its closest super class in the hierarchical taxon-
[ Figure 1: An illustration of our proposed hierarchical nov-
omy of known classes. To this end, we propose two different
elty detection task. In contrast to prior novelty detection
2 approaches termed top-down and flatten methods, and their
文档评论(0)