OLIE

OLIE was aimed to reconstruct the original image with the objects removed (editing) by incorporating the mask features as input to the reconstructor.

After performing many experiments, we scrapped the idea. However, some of the experiments that I performed were:

  • Using a Perturbator Model: Taking inspiration from GanPaint, I tried editing images using perturbations in the reconstructor, however it didn’t work well.

  • Using Gated Convolutions: Taking inspiration from Free-Form Image Inpainting with Gated Convolution, I used Gated Modules instead of standard modules which gave even poorer results.

  • Using Two Stage Model: Getting hole images from the reconstructor and then passing that into an inpainting model gave somewhat good results as the generating the hole image could be learned accurately to an extent, however, the results were no better than the common inpainting ones.

Our aim was to come up with a Single Stage Model for image editing using instance maps.

For more info, please see the Code on GitHub.

Jitesh Jain
Jitesh Jain
Ph.D. Student

I am currently working on developing multimodal vision systems. I am always eager to explore more AI research fields. Reach out if you want to collaborate!

Related