search

mihaidusmanu / d2-net

D2-Net: A Trainable CNN for Joint Description and Detection of Local Features
Other
761 stars 163 forks source link

I encounter a problem when I run on the Fountain,Herzjesu,southbuilding dataset use d2_pth.tf model #77

Closed fantuslk closed 3 years ago

fantuslk commented 3 years ago

a1 a2 the result of Fountain is normal butbut thethe result of the Herzjesu and southbuilding's dense points is too small... Please give me some advice.

mihaidusmanu commented 3 years ago

This is strange. Can you please post an image of the sparse model as returned by colmap? I suspect the poses are somehow wrong. I haven't evaluted D2 recently on these datasets so I am not sure where the issue might be coming from.

Also did you try other methods, e.g. SIFT, to make sure the data for these scenes is not corrupted?

fantuslk commented 3 years ago

Hi mihaidusmanu， Thank you for your reply. Can you give ,me the results of three datasets run on D2-Net? I can't reproduce them for the moment.

fantuslk commented 3 years ago

d2fcon1111 the sparse model result of foutain. d2hecon the sparse model result of herzjesu. I can't tell the difference between them.

fantuslk commented 3 years ago

Hi mihaidusmanu， I sorry.I found my mistake. I forgot to delete the previous reconstruction results, resulting in the dense reconstruction results that remain unchanged. Now other results are normal, that is, there is still a big difference between the mean track length and other methods.The paper uses the mean track length value as the track length or other values? d2herzjesu

mihaidusmanu commented 3 years ago

Hello. A shorter track length is normal. D2-Net has a lot of features (sometimes clustered together) and that ends up splitting longer tracks into multiple shorter ones. The following paper also reports similar results: https://arxiv.org/pdf/2003.08348.pdf