Contrastive learning of Class-agnostic Activation Map for Weakly Supervised Object Localization and

paper: Contrastive learning of Class-agnostic Activation Map for Weakly Supervised Object Localization and Semantic Segmentation

One sentence summary : Think of the prospect and background feature Natural similarity is low , And the front of the same texture / background feature High similarity , So directly in feature Then the segmentation head predicts the background score before , And feature Before multiplying and taking out / background feature. In a batch Before inner lowering - background pair The similarity , Improve prospects - Foreground and background - Background similarity , But according to the similarity ranking, reduce the pair Of loss The weight .

---

In detail , By reducing the dimension of features, we can find
（1） A foreground with a similar texture feature The similarity is high ,
（2） The similarity between foreground and background is low
（3） The background with similar texture has high similarity

therefore , In pre training encoder Extract feature after , Before direct prediction / Background score , And reinforce the above conclusion with comparative loss , So that the former / Background prediction is more accurate

As shown in the figure above ,$h(\cdot )$ yes Preliminary training Of encoder,$z_j$ It's the extracted feature map , adopt Random initialization Binary segmentation header $\phi (\cdot )$ Get random ( At the beginning of the ) Segmentation results , Such as fraction P
notes ： I don't know at this time P Does it represent the prospect or the background , Don't worry about... For the time being
At this point through P or (1-P) And feature Multiply , Can take out random ( At the beginning of the ) front / background feature.
In a batch Inside , Calculate the foreground - background pair The similarity , and Minimize the similarity ：

For the same kind （ Same as foreground 、 Background class ）feature, First calculate the similarity , ranking , Assign weight according to ranking ：

The higher the ranking ( The smaller it is ) Of pair The greater the weight of , The smaller the vice . This is because Not all backgrounds / The prospects are similar , Only those with similar textures are similar .

This means that I hope to begin with feature The above assumption is satisfied , In this way, we can find the similar texture directly through the similarity region, Then by comparing the losses, the score cuts the head to predict the prospect - background pair Small similarity （ This actually requires the segmentation head to accurately segment the foreground / background region?）, At the same time, it requires prospects - The foreground similarity is large , In fact, it is a requirement batch Inside cross-image Objects with similar textures are similar .

Counterexample
（1） Suppose a batch Inside all the foreground / The background textures are not similar
（2） Suppose a batch There are only a few sample Before / The background texture is similar
（3） Before hypothesis / There are multiple objects in the background , More complicated , The mixed texture of multiple objects still satisfies the assumption ？

Of course , In the end, this just completes the binary segmentation , For multi class goals , Can only assist CAM To do it ：