Image Segmentation mask aliasing

[YJonmo]

Describe the question.

Thanks for this work.

I have a pipeline for training image segmentation modes. I am using albumentations library for data augmentations. Now it is time for me to try DALI to get a speed boost.

I need to load the images and corresponding masks and train a model using augmented images and masks. However, I noticed when I Ioad the masks using the following lines the aliasing happens and new pixel values appear in the mask. In other words, the pixel value in the mask frame should be only a certain number such as 0, 25, 100, 220. But after operations such as fn.resize or fn.random_resized_crop values in between them appear.

    image_files, _= fn.readers.file(file_root=str(images_dir), file_filters='*.png', seed=1234, name="main_reader")#, num_shards=world_size, shard_id=global_rank)
    mask_files, _= fn.readers.file(file_root=str(mask_dir), seed=1234)

    images = fn.random_resized_crop(images, size = (512,512), random_area=[0.08, 1.0], 
                                    random_aspect_ratio=[0.75, 1.333333],antialias=True)

Before resize:

After resize:

So how can I avoid this?

Check for duplicates

• [X] I have searched the open bugs/issues and have found no duplicates for this bug report