Questions on deployment feasibility and zero-shot generalization

[mohamedsamirx]

First of all, congratulations on this impressive work.

I had two questions and would be very interested to hear your thoughts.

Do you think OccAny could be adapted for real-time deployment, or is it mainly intended for offline inference at this stage?
If real-time deployment is not currently feasible, do you think it would be practical to replace SAM3 with a lighter alternative such as EfficientSAM3, and replace Depth Anything with a TensorRT-optimized version?

Do you expect OccAny to generalize to more unstructured environments, such as a university campus, where there may be no clear roads or lanes, but instead buildings, open spaces, static obstacles, and dynamic objects?
In such a setting, do you think the method could work reasonably well zero-shot, or would fine-tuning likely be necessary?