Open Muhammad-Ibraheem-Siddiqui opened 1 month ago
Thanks for your attention. The pretrained weight of FSOL has uploaded along with how to load them:)
Thanks for the prompt response. While testing on FSC-147 using the provided weights, I am getting MAE = 18.6 and MSE = 104.6 in the one-shot case (screenshot attached). This does not coincide with the results stated in your paper i.e, MAE = 20.33 and MSE =68.28. Can you explain this change or is there something else I am missing? Thanks.
I just realized that it is the results on the evaluation dataset that you have reported in the paper for the FSC-147, one-shot case. I managed to replicate the results. Can you comment about the performance on the test set as mentioned in the comment above? Thanks.
Sorry for the late response. In our paper, we report the results on the evaluation set. In the ablation study, since our FSOL model is a localization model, we skip the counting performance and focus on the localization metrics. Your results are similar to our recently attached reproduction results.
Regarding the difference in counting performance between the validation set and the test set, please note that in the FSC-147 dataset, the categories in the validation set and the test set are disjoint, while the FSOL model does not require test-time adaptation. So, it is natural to expect differences.
Hi,
Thanks for the great work. Can you share the training checkpoints for FSC-147 dataset? I have tried to train on single GPU considering FSC-147 dataset for 200 epochs but the results are not as reported in the paper. Thanks in advance.