For the S3DIS, I observe that the MIOUS, AP@50 are very close to the final results in the initial epochs. But when I used the initial checkpoints for testing, the results are very bad! Do you know why it is showing good results for validation data but bad for test data?
For the S3DIS, I observe that the MIOUS, AP@50 are very close to the final results in the initial epochs. But when I used the initial checkpoints for testing, the results are very bad! Do you know why it is showing good results for validation data but bad for test data?