ZSL論文閱讀

1.記錄自己總結的規律，相信可以慢慢進入正軌。

2.爲了鍛鍊英文且不失去一些沒讀懂地方的原意，用英文記錄

目錄

Learning To Detect Unseen Object Classes by Between-Class Attribute Transfer

Zero-shot classification by transferring knowledge and preserving data structure

Zero-shot learning via discriminative representation extraction

---

1. Learning To Detect Unseen Object Classes by Between-Class Attribute Transfer

   1. Method：Based attribute
      • Learning a attribute from a range of different classes
   2. Using:A attribute layer
   3. Problem:Need a lot of manually labeled training data
   4. Problem setting:
      • Problem：learning with disjoint training and test classes
   5. Two method：
      • DAP：uses an in between layer of attribute variables to decouple the images from the layer of labels
        1. Using the MAP(maximum a posterior) to predict unseen class
        2. 在實現時，爲什麼可以忽視p(z)
        3. p(z|x)=p(z|a)p(a|x)是什麼公式
      • IAP:the attribute is from the connection layer between the classes are known or unknown
   6. Result
      • DAP perform better than IAP

2. Zero-shot classification by transferring knowledge and preserving data structure

   1. Method:
      • transferring knowledge and preserving data structure
      • New Points:
        1. Transferred knowledge from source domain, the target classification model directly connects images and labels
        2. Use the attribute directly to learn semantic correction rather as a middle layer
        3. Preserving manifold data structure to handle domain shift problem
        4. ONE stage and solving the domain shift problem
   2. Manifold assumption
      • If the samples are close in the low_dimension,they are also close in the origin space
   3. Like SAE

3. Zero-shot learning via discriminative representation extraction

   1. Problem:
      • Visual feature is not the best choice for zero-shot-learning because that it is for multi-shot task
      • While the number of images more than classes,the mapping that maps all samples to one semantic space will not perform well
      • Due to the hubness problem,inferring a test image in visual feature instead semantic space is not perform well;
   2. Point
      • By discriminative zero shot visual representation
      • Add a preprocessing stage before training stage
      • Preprocessing stage:
        1. Reduce the dimension of the visual feature in order to get the template of each class
        2. Let the space be representation space
      • Training stage
        1. Using the semantic embedding to learn a class template by a nonlinear regression.
      • Testing stage
        1. First reduce the dimension
        2. Compare the similarity to decided which class belong to.
      • Existing ZSL methods are based similarity and based projection
      • Methods based on compatibility:
        1. Advantage: Take into account of the large margin
        2. Disadvantage: Don not make use of the neighborhood information
      • Methods based on similarity:
        1. Advantage: Take use of the rich information of the seen class
        2. Do not meet the key factor to improve the accuracy
      • Methods based on matrix:
        1. Can be seen as linear methods and also keep the information while transfer the information
        2. May ignore the nonlinear character reside on the dataset
   3. Contributions:
      • Verify the supervised discriminant training on seen class can benefit the unseen class
      • Aggregated perform better than Large Margin on ZSL
      • Advantage:
        1. Directly inferring class in the visual representation space