Dali: INTERP_LANCZOS3 doesn't work for downsampling

[ei-grad]

This pipeline results in images looking like images downsampled with INTERP_NN:

from nvidia.dali.pipeline import Pipeline
import nvidia.dali.ops as ops
import nvidia.dali.types as types


class LanczosPipeline(Pipeline):

    def __init__(self, batch_size, num_threads, device_id, file_root):
        super().__init__(batch_size, num_threads, device_id)
        self.input = ops.FileReader(file_root=file_root)
        self.decoder = ops.ImageDecoder(device='mixed')
        self.resize = ops.Resize(
            device='gpu',
            resize_longer=224,
            interp_type=types.DALIInterpType.INTERP_LANCZOS3,
        ) 

    def define_graph(self):
        images, labels = self.input()
        images = self.decoder(images)
        images = self.resize(images)
        return images

dali=0.22.0 installed with pip

[mzient]

We use Lanczos filter as an interpolation filter, not a low-pass filter, and therefore the behavior is more or less expected. Current solution would be to use a different filter (e.g. Gaussian) for downscaling than upscaling, which can be done with:
min_flter=types.INTERP_GAUSSIAN, mag_filter=types.INTERP_LANCZOS3.
We can think about changing how INTERP_LANCZOS3 behaves when downscaling, since now it's admittedly not very useful for that purpose.

[ei-grad]

We use PIL.Image resize with Image.LANCZOS (==Image.ANTIALIAS) interpolation type, and it gives a way better results applying to image features calculation. With INTERP_GAUSSIAN the edges in downsampled image are significantly blurried, resulting images are visually less detailed than images resized by PIL with lanczos.

[mzient]

We'll look into it, but the earliest possible release with Lanczos downsampling fixed is 0.24, if we scope it for the release.

[mzient]

@ei-grad The fix is already on master branch, it should be there in release 0.24 and, of course, next nightly build.

[JanuszL]

0.24 has been released. Please check if that works for you.

[ei-grad]

Thanks! We now using DALI for inference, works like a charm. Images are visually identical with PIL, though it gives some difference on pixel values.

[JanuszL]

@ei-grad - we are happy that it works as you like!