Dustbin and Homography Adaptation

[johnnyboloni]

Hello everyone,
I'm implementing my own training code for SuperPoint and was wondering what to do with the dustbin when doing Homography adaptation.
To implement Homography Adaptation, I need to convert the 65 detection channels back to spatial pixels, unwarp the resulting heatmap for each sampled homography, and sum resulting heatmaps over all sampled homographies. Of course the problem is the extra dustbin channel - it has no "spatial" value to it. I have a couple of ideas how to deal with this:

1. Weight pixel detections by inverse dustbin value (this way, pixels coming from a grid cell with a strong dustbin get a very low weight)
2. Apply dustbin logic before unwarping heatmap

Any idea? What did you guys do?

[rpautrat]

Hi,

We used option 2.