Hacker News Reader

What are the current limitations of the model which you're working on? Could you share some examples where the output is currently sub-optimal and what steps you're taking towards improving it?

View on HN

Hi Alex, thanks for the question! Currently, we are not producing great results when it comes to realistic humans, animals, and vegetation. Those requests are not served as we haven't yet trained for these kind of inputs. Once our geometry reconstruction has high accuracy for hard-surface objects we'll move on to this category, collect data, train and improve.

> Once our geometry reconstruction has high accuracy for hard-surface objects we'll move on to this category, collect data, train and improve.

is that really possible though? For complex objects there should be multitude of 3d structure that fit a 2d projection and there's really no way to say "which one is the most correct one".

Thanks ekianjo! Yes, that is true. For complex objects, the effect of the algorithm is to lose detail. We mitigate this with our multi-view option (up to 6 images of different views). At the moment high level of complexity is in most cases not processed. We do best with simple objects.

Reply on news.ycombinator.com