pretrained models with fixed bbox masks about edge-connect HOT 3 CLOSED

@Rayna-Zhang We are updating the model and will release weights that account for both regular and irregular masks.

Stay tuned!

from edge-connect.

@knazeri thanks for your Excellent work！ When I use your pretrain model on places2 to initial the network and then we train the model with model 3 on mask 1（random block）。the test result seems not good？ is there any tricks for random block？

from edge-connect.

@ljjcoder Random blocks are a very challenging case of inpainting, the reason being is that 25% of the entire image is missing and we need to fill that with a perceptual meaningful content. If you look at most state of the art models out there, their results all suffer from some artifacts when a large portion of the image is missing! For the record, in our model, it is the first stage of the inpainting process that needs improvement! You can test it by setting MODEL:2 and see the quality of the inpaint model.
There are ways to improve the quality of the edge generator. For one, we found GAN hinge loss to outperform nsgan version! You can also start training on lower resolution images (Say 128x128) and use the weights for higher resolutions! There are other techniques that we are also working on right now and will publish our findings soon.

from edge-connect.