Which loss did you? dense or sparse?

Hi, there

I find it is confusing that there are dense loss and sparse loss. The config you provided used dense loss function, while the original paper used sparse loss.

Which one did you use for the provided weight?

By the way, what does heat2.0 mean in model name?

Cheers

Hi @YanShuo1992,

Thank you for your question.
I implemented dense loss then sparse loss. They have similar performance in my experiments.
The COCO model used sparse loss (

pytorch-superpoint/logs/superpoint_coco_heat2_0/config.yml

Line 84 in 4ff74df

enable: true

).

For heat, I named when I was experimenting adding gaussian kernel to the labels, i.e. heatmap.
(

pytorch-superpoint/logs/superpoint_coco_heat2_0/config.yml

Line 40 in 4ff74df

gaussian_label:

)

Hope this helps.
Best,