Hello, thank you very much for open-sourcing your code. But I have some questions about dataset splits. I noticed that you wrote in the readme that the division of your tasks is slightly different from ORE.
I see that you have replaced VOC with animals etc. for task1. and the number of images is different for several tasks after that. May I ask why you are doing this?
Also I have a question about the way the data is sliced for both. As per the paper, task1 is trained with VOC as known data. The later ones are tested as unknown data. But when task2 is done, the training is supposed to be on images of categories like VOC and outdoor. Then why the train images are only 50,000, rather smaller than the 80,000 of task1.
And also does test image refer to the test in the process of training a model with known classes? If so. Then is the test image for the unknown class using all the test and train images for the later classes?
Hello, thank you very much for open-sourcing your code. But I have some questions about dataset splits. I noticed that you wrote in the readme that the division of your tasks is slightly different from ORE.
I see that you have replaced VOC with animals etc. for task1. and the number of images is different for several tasks after that. May I ask why you are doing this?
Also I have a question about the way the data is sliced for both. As per the paper, task1 is trained with VOC as known data. The later ones are tested as unknown data. But when task2 is done, the training is supposed to be on images of categories like VOC and outdoor. Then why the train images are only 50,000, rather smaller than the 80,000 of task1.
And also does test image refer to the test in the process of training a model with known classes? If so. Then is the test image for the unknown class using all the test and train images for the later classes?
Thank you! I look forward to hearing from you.