VOC dataset preparation difficulties about blitznet

[dvornikita]

Hi,

The reason for this assert is that there are no segmentation annotations for the VOC07 (in fact, there are, but we didn't use them for training at all, treating it as a detection only dataset). If the segmentation flag is off, we still create segmentation masks for images but fill them with zeros. This is done to preserve the structure of the data independent of the dataset. However, you are right that there should be segmentation=segmentation, in line 210 of datasets.py. Thanks for noticing this, I'll push that.

The numbers that you see in splits_to_sizes reflect the sizes of /ImageSets/Main/train.txt rather than /ImageSets/Segmentation/train.txt. If you would like to include the VOC07 segmentation data in the training, you need to delete the assert you mentioned before and modify the function read_segmentation(...) in voc_loader.py to create void segmentation mask if the gt mask wasn't found in the folder (that's what I would do).

Thank you very much for the meaningful feedback. If you have any other questions, contact me.

[revilokeb]

@dvornikita thank you for your reply, got it!