Performance gap between self-tested and on paper about softgroup HOT 7 CLOSED

Awesome! Saved my day! Congrats again for your great work!

from softgroup.

Sorry that I pasted this twice. I ran the inference command and get this result.

./tools/dist_test.sh softgroup_scannet.yaml softgroup_scannet_spconv2.pth 1

Am I getting this correctly?

from softgroup.

Yes. This is the performance difference when we convert checkpoint from spconv1.0 to spconv2.1. If you want to use paper results. You can use the legacy branch.

from softgroup.

So the weights is permuted directed from the spconv1.0? So if I retrain this spconv2 with dataset scannet, should I get similar results as the in the paper? Do you have any idea of what's the cause of this huge drop?

from softgroup.

I think 0.3 point difference is acceptable. I think you may be confused between reported results (maxAP=100) and the code (maxAP = 1)

from softgroup.

Oh, I see the difference. That's the hidden test on the benchmark. Thank you so much for you kindly reply. Actually I am also working on migrating my pretrained model from spconv1 to spconv2. Is that possible that you share your code? That will really help me a lot.

from softgroup.

Please check tools/convert_checkpoint.py for the details.

from softgroup.