Runtime and Evaluation

[gauenk]

Hello Team 👋

Thank you for releasing your code. It has been a breeze to install and work with. I love the core idea of your project, and I'm doing research in a complementary area. I think the idea of densely processing sparse regions is really smart, and as you've shown, it's quite successful. I have two questions:

1. I was curious if there was a reason why you report runtime per-object rather than per-image. I think runtimes of segmentation networks are usually reported per-image, so the comparison in your paper didn't see quite fair. I may be mistaken, so any comments would be appreciated.
2. I saw you use the ground-truth location and spatial covariance term in your network at test time. After some thinking, I suspect these summary statistics are analogous to "prompting" from LLMs. Did you/your team try estimating mean+shape with a network, and if not, why you did not do this? One challenge I can think of is that small objects are labeled inconsistently. This would make training/evaluation a nightmare. I wondered if this was part of your rational and/or if something else was a factor here.

Thank you.

Best,
Kent