Closed Liming-belief closed 7 months ago
This could be expected since the model is only trained with low-resolution inputs. One workaround is to do the inference at resized low resolution images and finally resize the predicted flow to the original resolution.
I predicted the optical flow very well at a resolution of 960x520, but the prediction effect was very poor at a resolution of 1920x1080. Do you have any solutions。 Thanks