skingorz / fd-align

Hi there 👋

fd-align's Issues

Questions about the results in the paper

Hi, thanks for your great work! I have a few questions regarding your paper.

1. Do the results of the FD-Align used in your tables/figures (Table 1, 3, and Figure 5, 10) include SPC? (Since you only emphasized in Table 2 that FD-Align with SPC)
   I think the answer is yes.
2. Which dataset was used for ablation, and which visual backbone was employed? (in Table 4)
   I assume it was ImageNet with ViT-B/32, but I have some additional questions based on this assumption:
   • The first line of the CLIP zero-shot result is 60.33, which is similar to the result obtained by using ResNet50 as the visual backbone with 7 prompts (in Tip-Adapter).
   • The second line of Fully Fine-tuning results seem to be the average results over 11 datasets shown in Figure 10, but ZS CLIP and FD-Align ablation results are not matchable if so. (By the way, the average results of FT are different in Figure 5 and Figure 10)
   • The results in the subsequent lines seem to be higher than expected for ViT-B/32, suggesting that ViT-B/16 might have been used instead?