Loading VOC 2012 dataset with Dataloaders

vision
1

VOC 2012 dataset consists of images and their corresponding segmentation maps. I want to apply similar transforms to both the image and its segmentation map while loading. Any suggestions about how to proceed for this task?

1 Like
2

We don’t have a ready solution implemented, but there has been some discussion in a torchvision issue.

2 Likes
3

@Gaurav_Pandey you can easily adapt a dataset to handle co_transforms in the __call__ function (e.g. see this gist which has general structure that handles co_transforms):

and here are some relevant affine transforms to actually use – you’ll see the transforms must take in two arguments for the input and target images:

3 Likes
4

Thanks guys. That was very helpful.

5

I also implemented a dataset for VOC2012.

6

@apaszke does pytorch have any plan to add VOC dataset like what it has for coco?

7

Now Torchvision has VOC dataset implemented, but it’s not that user friendly. Specifically, it doesn’t allow co-transform (data augmentation such as rotation and scaling) and separate transforms (normalization on images) to exist at the same time (raise an error here). Many good third-party implementations like this one.