Appendix A Further Implementation Details

In the main manuscript, we have provided the most important part of training and testing details. Here, we add other details about the training procedure, baseline implementations, etc. We first remove the stem module and then repeat level embeddings to the same shape as visual prompts. The repeated level embeddings are treated as visual prompts. To analyze the impact of the proposed adaptive pseudo-label correction strategy, we conduct experiments on two model variants. Thus, We set γ = 0 .65 for all experiments.