How to repeat Figure 10 results in the paper?

[knazeri]

How to repeat Figure 10 results in the paper?

[knazeri]

@jtetc Yes that's exactly how you can test it but, the pre-trained weights that we released are not good to reproduce entire right or left portion of the image. What we did for Figure 10 was to train the model on a half-region mask MASK: 2 only on G2 MODEL: 2 and then write an script to merge two image edge-maps before running the evaluation.

[dulibubai]

Hi! As you mentioned in your paper ,I want to know what is the fixed mask dataset? Are they the 12000 test mask dataset made by Liu Guilin in PartConv?

[knazeri]

@dulibubai This is from our paper:

Regular masks are square masks of fixed size (25% of total image pixels) centered at a random location within the image

Basically, they are fixed size square masks that you can use by setting MASK: 3 in the configuration file.

PS. we randomly create fixed size masks during training and there's no need for an external mask dataset.