Why using Gaussian Filter in MDA ? about dirl-inharmonious-region-localization HOT 3 CLOSED

Thanks for your attention!

1. Gaussian Filter is expected to enlarge the attention region to make final mask prediction more robust.(In some extreme cases, one can observe that attention region might cover few part of ground truth, yielding a poor quality mask). However, due to limited time, we conduct very few experiments, and found that the improvement is imperceptible when the model is equipped with Gaussian Filter. We will clean the Gaussian Filter code in future.
2. I forgot to revise the sentence. Resizing ground-truth(GT) mask M to A_{k} will loss clear boundary when the target resolution is too small(e.g., 256x256->16x16), and the resized GT mask will be empty in some extreme cases. These resized GTs will lead to wrong guidance and hurt performance. Resizing A_{k} to match the GT resolution is much better, since interpolation operation in mask prediction is differentiable.

from dirl-inharmonious-region-localization.

Thanks for your feedback!
I also found in your code that using iou_loss in bce_ssim_loss, which is not mentioned in your paper.
I wonder whether iou_loss can greatly improve the final result.

from dirl-inharmonious-region-localization.

Thanks for your feedback!
I also found in your code that using iou_loss in bce_ssim_loss, which is not mentioned in your paper.
I wonder whether iou_loss can greatly improve the final result.

iou_loss is aim to improve the IoU of the final inharmonious region prediction, and it indeed improves IoU score and slightly improves other two metrics.

from dirl-inharmonious-region-localization.