How to use customized segmentation mask for test?

[Lotayou]

I want to replace the OneFormer used in your demo with some other segmentation models (for instance, SAM), now suppose I have

input_img: H1 * W1 * 3, uint8
input_mask H1 * W1, binary 0,1
ref_img: H2 * W2 * 3, uint8
ref_mask: H2 * W2 binary 0,1

How should I modify the demo code to get correct structure and appearance guidance features? Thanks for your help in advance!

[Lotayou]

Also，what's the expected dtype, shape and data range for img and mask here? I suppose the img should be torch.float32 and [-1，1], but what about the mask input? Can I directly use a binary 0, 1 mask?

[vidit98]

Yes the image should be torch.float32 and [-1，1] and mask should be of type int(), so it can be binary as well

[CesarERamosMedina]

@Lotayou btw, if I understand correctly, to run this with a different segmentation model you are not only going to need a binary mask, but also the panoptic and semantic segmentation masks
If I understand correctly, you pass to get_appearance: {the input image and the panoptic mask (which isn't binary)} and {the reference image and the reference mask (which is binary)} , but please correct me @vidit98 if wrong

[vidit98]

Yes, that is correct and if a different segmentation model is used the class mapping should match that of the trained model else a new model has to be trained. The pipeline @CesarERamosMedina is talking about is for the demo, however, I see the picture above from the model definition.

[CesarERamosMedina]

I actually meant from the model itself. I think the code in your gradio app makes two calls to that function, one with a binary mask and one with a non-binary mask