Open llj110 opened 2 months ago
Thank you for your message. We made slight adjustments to the training parameters on other COD datasets, increasing the batch size from 4 to 8. This change was made to fully utilize the GPU memory since there were no depth maps in the other datasets. Similarly, the loading method for these COD datasets was implemented by removing the depth information loading in dataloader.py. I hope this clarifies your question.
Thanks for your valuable reply. Here, I have an extra question: For the dataset of NC4K, at what ratio did you divide the training and testing sets?
The entire NC4K dataset is treated as part of the test set.
Do you mean to combine the training set of COD10K and CAMO to train a model, and then test it separately on the testing sets of COD10K and CAMO, as well as NC4K?
Yes, that's it.
Is the training parameters of other COD datasets consistent with those of ACOD-12K? Can you provide the loading method and training parameters for these COD datasets?