search

mihaidusmanu / d2-net

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

Performance of trained model weights #21

Closed a1302z closed 4 years ago

a1302z commented 4 years ago

Hi, thanks for making your code public. I was wondering if there are performance stats to the trained model weights you offer? Would be great for deciding which weights to use for experiments.

mihaidusmanu commented 4 years ago

Hello.

d2_ots.pth are simply the off-the-shelf Caffe weights pretrained on ImageNet. By default, at the moment, you should use the model fine-tuned on the entire MegaDepth - d2_tf.pth (i.e. the one used by the feature extraction - extract_features.py - script with default parameters). The model fine-tuned without the PhotoTourism scenes was only released in order to have a fair comparison on the Image Matching Workshop Benchmark - https://image-matching-workshop.github.io/. We plan on releasing new models in the future and we will include a more detailed comparison at that point in time.

a1302z commented 4 years ago

Thanks for the detailed answer.