ArantxaCasanova
commented
3 years ago Hello,
it does seem like the unseen images are quite similar to the training images, so I would guess IC-GAN could work well for this kind of task. However, it might be useful to do an analysis of which feature extractor would better preserve the main object's features (as opposed to capturing the entire scene's features), to make sure you preserve that content. Alternatively, one could add additional constraints during training to make sure the main object is always preserved with high quality in the output generations.
I hope that helps,
Arantxa
Hi Community, I am trying to solve an image generation problem using IC_GAN, but I am not sure if IC_GAN is the right direction to move forward or not and therefore I am posting on this forum. I have attached a link to google drive. The first slide contains the training data and second slide contains the test data or data we want to generate.
Link : https://docs.google.com/presentation/d/1V3DysdfP4lJQRgAu9m8GDBnH17PKoxQU/edit?usp=sharing&ouid=108571067697385919652&rtpof=true&sd=true
Objective is to generate new images (new product) in different orientation and random backgrounds. Given that the new image (new product) is not directly trained in the model, but similar products are used to train the model. I have tried XingGAN and pose-gan for this problem but they were not successful.
Thank you.