Image embedding space - Fig. 6

[DrMachado]

Hello and thank you for this great work ! It is unclear to me how to extract the CLIP label encoding for downstream tasks as shown in Fig. 6 of the manuscript, with the t-SNE plot of the embedding space. Thank you for your time.

[ljwztc]

After obtaining the feature map, you can scale up to the input size. Then you can find the feature vector for each category with ground truth.