search

munanning / MADA

Multi-Anchor Active Domain Adaptation for Semantic Segmentation (ICCV 2021 Oral)
MIT License
43 stars 4 forks source link

If a image not containes all the categories, how to 'connecting the vectors of all categories into a long vector'? #2

Closed luyvlei closed 3 years ago

luyvlei commented 3 years ago

As I understand, MADA need to generate a image-level featurte vector with the mean vector of all classes in a image. But if a category is missing in a picture(in fact it is a common phenomenon), how to generate the vector?

BinhuiXie commented 3 years ago

As I understand, MADA need to generate a image-level featurte vector with the mean vector of all classes in a image. But if a category is missing in a picture(in fact it is a common phenomenon), how to generate the vector?

In my opinion, you can fill with zeros for absent classes in an image.

luyvlei commented 3 years ago

It should be the case. Then the k-means will cluster the images with the same category(or missing the same category) into a cluster.