mhamilton723 / STEGO

Unsupervised Semantic Segmentation by Distilling Feature Correspondences
MIT License
711 stars 142 forks source link

Fine-tuned Model #67

Closed Holmes2002 closed 1 year ago

Holmes2002 commented 1 year ago

Many issue when using cluster algorithm, such as backgound same color with our labels , brightness difference of the same label, ... I think we should use Contrastive to learn feature of total images. Then add classifiers and train with labeled dataset. Is any one do this and answer for me ?

mhamilton723 commented 1 year ago

@Holmes2002 im ot quite sure what you mean here. Is there something you think the cluster probe is not capturing in your data? Try varying the number of clusters to see if your dataset naturally has more or less. Also iyou could clarify your question that would be a big help!

Holmes2002 commented 1 year ago

I solved it. Tks for your cmt !