search

ZhangGongjie / Meta-DETR

[T-PAMI 2022] Meta-DETR for Few-Shot Object Detection: Official PyTorch Implementation
MIT License
388 stars 84 forks source link

Some questions about t-SNE #67

Closed Jackman-Xu closed 1 year ago

Jackman-Xu commented 1 year ago

Thanks for your great and inspiring work! The designed mechanism of attention enlightens me.

However, I still have some questions about the t-SNE visualization when I read the paper and codes. I wonder if the t-SNE visualization was made by using: (1) the category codes' features obtained by RoIAlign and GAP, or (2) the decoder's final output features (the correct predicted ones)

image image

Jackman-Xu commented 1 year ago

Probably the author is using the support class prototypes, which calculated by RoIAlign & GAP using the GT-bbox, to perform t-SNE. I seem to have reproduced the t-SNE experiment under the 10-shot setup, but not 2-shot setup.