search

ltttpku / ADA-CM

19 stars 2 forks source link

The criteria for selecting instances #11

Closed slcheng97 closed 3 months ago

slcheng97 commented 4 months ago

I'm intrigued by your method of extracting human/subject/union ROI features. Specifically, I'm curious about your approach to determining the number of instances to store in the concept memory, as well as the criteria for selecting instances to include or exclude from the concept memory. I also found tha same issue in #4 . Would it be possible for you to share the code you used for this purpose?

ltttpku commented 3 months ago

Hi, I don't plan to release the code for extracting features from human, subject, or union regions of interest (ROI). The features are extracted using ROI pooling on the feature map generated by CLIP, utilizing bounding boxes provided by DETR. This process should not be too difficult to implement. For the strategy of constructing concept-guided memory, you can refer to our supplemantary material for more details.