Inference on other parametric models

[raincrash]

Hi, thanks for the great publication. A quick question on using a-csm on other custom parametric models. For inferencing on categories not present in CUBS, is there a way to use the given template models and annotations?

[nileshkulkarni]

Hi @raincrash,

I'm slightly confused about your question. CUBS only has bird images. Could be a bit more specific so that I can answer your question better.

If the question is that you want to infer on images, not the CUBS dataset?
yes, it is possible to do it. You can modify by the data loader slightly, and use the predictions from the model.

For inference, the model only requires a center crop image of a single object of size 256 x 256. I think this function might be useful here. You could refer a call to it here

Thanks,
Nilesh

[raincrash]

Hi, @nileshkulkarni sorry for the late reply. Yes, I wanted to infer on images that are not on the CUBS dataset, but also on a different parametric model, for eg airplanes. Do I need to train the network for each type of parametric model (i.e create a new parameterization of the surface to a new template mesh, train the network from scratch to learn the CSM), or is there was a way to infer on unseen categories (e.g. airplane images and model) directly from the existing pre-trained weights?

[nileshkulkarni]

Hi @raincrash,
Ours is category-specific model so it will not work on categories it is not trained on. To train on a new category you would have to create a first a category-specific parameterization and then train CSM or ACSM style model to infer the surface parameterization. It would incorrect to run inference on airplanes while using weights from a model that is trained on birds.

[raincrash]

Got it! Thanks. I'll close the issue.