论文中值得摘抄的句子
# 论文中值得摘抄的句子
- We abandon the memory queue, which we find has diminishing gain if the batch is sufficiently large:我们抛弃了内存队列机制,我们发现当批样本量相当大的时候,内存队列的性能增益会减小
- which indicates the misuse of the weak labels:这说明弱标签的使用不当
- the proposed dual branches separately handle full and weaksupervised learning and effectively eliminate their mutual interference:所提出的双分支独立的处理全监督和弱监督学习,能够有效消除他们的彼此干扰
- Besides, separate training prohibits the exchange for supervision information:此外,单独的训练组织了监督信息的交换
- Our DDF method facilitates the feature disentanglement at the global and local stages:我们的DDF方法促进了全局和局部阶段的特征解构
- To address this dilemma:为了解决这一难题
- To alleviate this issue:为了缓解这个问题
- Motivated by the aforementioned observations:出于上述观察的动机
- The deficiency of segmentation labels is one of the main obstacles to semantic segmentation:分割标签的不足是语义分割的主要障碍之一
- The characterization of similarity and discrepancy is at thecore of all clustering methods. :相似性和差异性的特征是所有聚类方法的核心
- The basic idea of SPICE is to synergize the discrepancy among semantic clusters, the similarity among instance samples, and the semantic con-sistency of local samples in an embedding space to optimize the clustering network in a semantically-driven paradigm.:SPICE 的基本idea是同步:语义类之间的差异性,实例样本间的相似度以及嵌入空间的局部样本间的语义一致性,利用语义驱动的范式来优化聚类网络
- However, CAM usually onlyidentifies the most discriminative object extents,which is attributed to the fact that the networkdoesn’t need to discover the integral object to recognize image-level labels.:然而,CAM通常只识别出最具鉴别力的物体范围,这是因为网络不需要发现物体全貌来识别图像级别的标签。
# 每日一句
- Dream big and dare to fail.
- 敢于梦想并敢于失败
上次更新: 2021/10/17, 16:23:18
- 02
- README 美化05-20
- 03
- 常见 Tricks 代码片段05-12