2020-01-03 CVPR 2020 Self-supervised Domain-aware Generative Network for Generalized Zero-shot Learn

CVPR 2020 Self-supervised Domain-aware Generative Network for Generalized Zero-shot Learning

摘要

引入自监督学习对测试数据进行域判别

动机

通过自学习将两个域的样本分开。

方法-SDGN

我现在理解呢，测试的时候，肯定是用的test_seen和test_unseen。训练的时候，可能挑出了一部分test_unseen。也就是把test_unseen分成两部分。一部分用于训练，一部分和test_seen混合作为测试数据。

3.3. Cross-Domain Feature Generating Module (CDFGM)

3.4. Self-supervised Learning Module (SLM)

Reference Anchor Learning.
We notice that each class can be represented as a attribute vector, where each dimension encodes a high-level visual property. This gives rise to the idea of training an attribute classifier and extracting its parameters as anchors.

For a series of attribute classifiers $\{g_1,g_2,\cdot \cdot \cdot ,g_{d_e}\}$ w.r.t. $d_e$ attributes, we extract their weight parameters as anchors: $\{A_1,A_2,\cdot \cdot \cdot ,A_{d_e}\}$, $A_i\in R_{d_v}$ .
$A_i$相当于是视觉空间中的每个属性的映射中心。

Domain Feature Reconstruction.
多属性标签
$${\mathcal{M}}_i^{(k)}=\frac{exp(<A_k,{\tilde{x}}_i>)}{{\sum }_{j}exp(<A_j,{\tilde{x}}_i>)}$$

Cross-Domain Triplet Mining.

3.5. Model Training and Prediction

总损失：

最后分类是用生成的目标样本和源样本训练一个分类网络。

3.6. Discussions

three relevant methods including CEWGANOD [22], f-VAEGAN[45] and SABR-T [27]
(1)22需要额外的分类器
(2)对比45和27，target discriminator用了多标签，类似于源域的监督信息。
(3)引入了自监督

实验

对比实验

消融实验

疑问

广义零样本下怎么使用测试数据？