So I'm trying to follow a paper that uses a AE to learn point clouds. The thing is, the dimension of the point cloud data is 3 (x, y, z), but the dimension of the latent space from what I can tell is 128.

I've always read that the point of a AE is to introduce a bottleneck for the latent space so that this space becomes a meaningful representation of the data. If the dimension of the latent space is higher than the dimension of the input data, how can the model learn properly. Is it not essentially just remembering the data instead of properly reconstructing it? Or is there some kind of difference that happens with an highly unstructured kind of data like point clouds?

Should I try switching the number of points for the location data, so that the dimension goes from (2000, 3) -> (3, 2000). This would enable me to work with a higher dimensional-space that is appropriate for an AE, but for some reason this idea does appeal to me. Any ideas?

This is the paper for reference: https://arxiv.org/pdf/1707.02392.pdf And this is some code that follows the paper: https://github.com/TropComplique/point-cloud-autoencoder



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.