Open bluesky314 opened 4 years ago
It is trained on inputs of size 320x320 but these also include larger size patches which have been downsampled to 320x320. The model performs well on images of 2000x2000. You are right though that it is not invariant to image size, and I suspect images of much larger resolution, say above 10 megapixels, would not work well
As the model is trained on small patches, will it have trouble inferring on large 2000X2000 images as it has never see details of this larger size? Don't all matting models suffer from this as all are trained on 320X320?