Questions on quantitative indicators APLS and TOPO

[Sysilia]

I ran the project with my own 512*512 sized dataset and got pretty good visualization results, but when I was quantifying the results I referred to spacenet's evaluation metric to quantify my own results, and I got topo results where a large portion of the prec was 1.0 and a lot of results close to 1, which I thought was very unreasonable, but didn't I don't know what's wrong with it, do you need to adjust any parameter when you quantize it?

[TCXX]

Thank you for trying SAM-Road!

Yes, I think you need to adjust params for your 512x512 dataset-

Buffer Size: The TOPO metric uses a buffer around roads to determine matches. For 512x512 images, try reducing the buffer size to match your image scale; start with a buffer size of ~2-3x your road width in pixels.

Matching Parameters: Adjust the max_path_length, min_length_px, and matching_threshold

It would be helpful if you can provide:

• The current buffer_size_meters you're using
• Your road width in pixels
• A sample of your evaluation configuration

[htcr]

I ran the project with my own 512*512 sized dataset and got pretty good visualization results, but when I was quantifying the results I referred to spacenet's evaluation metric to quantify my own results, and I got topo results where a large portion of the prec was 1.0 and a lot of results close to 1, which I thought was very unreasonable, but didn't I don't know what's wrong with it, do you need to adjust any parameter when you quantize it?

TOPO is essntially sampling subgraph patches from predictions & GT and comparing them. If prediction is accurate, you might indeed get a bunch of 1.0 (full score). Were you looking at the intermediate outputs or the aggregated final score?