it seems you are training and validating on the same dataset?

[sunwayforever]

although the ids is split into train and val, you are always sampling from the whole ids

[AkshayPS12]

Hi @sunwayforever if you see in voc_data.py in the generate function , the indices of train and val data are separated and they are being used for "sampling".

[showyamessage]

Yes, they are separated. But get_annotation and get_image functions get their filenames from self.ids. I think, transfering the filename (calculated in generate anyway) directly to get_annotation and get_image should solve this issue.