Running on different size images?

[ortegatron]

Hi, thanks for your paper and code. I'm trying to make it run on bigger images, kind of struggling with it. Do you think it will be easy or needs major code and model changes? Thanks

[HongwenZhang]

Hi, what do you mean by bigger images? More than one person in the image with higher resolutions? If so, you might need a person detector to crop person regions before using our model.

[ortegatron]

still one person by image, but on a higher resolution. For example, 500x500 pixels. Will it increase the accuracy, or would it be the same? I have just figured out how to make it work on bigger images, but the smpl output is just wrong.

[HongwenZhang]

In this case, you need to re-train the model with higher resolutions of input images and IUV maps. When using IUV maps with higher resolutions, the reconstruction performance should be better.

[ortegatron]

Awesome, thank you. Can you tell me on which platform did you train and how long did it take? Thanks again, I'm writing some nice stuff with it, will come back to you to let you know when its released!

[HongwenZhang]

Our model is trained on a single TITAN Xp GPU, which takes about one day. Feel free to let me know if you encounter problems or have any results.

[sunjc0306]

Hello. I would also like to run in higher resolution. I need to change the focal length. Do I also have to retrain it?

[HongwenZhang]

Hello. I would also like to run in higher resolution. I need to change the focal length. Do I also have to retrain it?

Hi, retraining the model should have additional performance gains.