search

XuyangBai / D3Feat

[TensorFlow] Official implementation of CVPR'20 oral paper - D3Feat: Joint Learning of Dense Detection and Description of 3D Local Features https://arxiv.org/abs/2003.03164
MIT License
261 stars 38 forks source link

Could you publish a pytorch implementation ? #13

Closed qxh123456789 closed 4 years ago

qxh123456789 commented 4 years ago

Dear XuyangBai,

Thank you for your great work! I have noticed that you have an implementation of KPConv. Have you done your pytorch implementation of D3Feat based on that ?

Best regards, Xianghua Qu

XuyangBai commented 4 years ago

Hi Xianghua,

Thank for your interest. Yes, I have implemented a PyTorch version of D3Feat but need some time to polish the code. I will release it as soon as possible.

Best, Xuyang

qxh123456789 commented 4 years ago

Hey Xuyang,

Thanks for your time. Btw, have you looked at Hugues' (official) pytorch implementation of KPConv ? Did you use that for an implementation?

Best regards, Xinghua

XuyangBai commented 4 years ago

Hi @qxh123456789

Yes, I implemented the D3Feat PyTorch version based on the official KPConv-PyTorch with some modifications.

qxh123456789 commented 4 years ago

Thank you @XuyangBai ,

I am looking forward to your work.

XuyangBai commented 4 years ago

Hi @qxh123456789 I just release the PyTorch code here https://github.com/XuyangBai/D3Feat.pytorch but current performance is a little bit lower than TensorFlow version. Please feel free to have a try.

qxh123456789 commented 4 years ago

@XuyangBai Thank you for your efforts. I will try this later. Btw, may I ask about its current performance ? Did you save the registration recall and FMR result of your current implementation ?

XuyangBai commented 4 years ago

Hi @qxh123456789 the current implementation can achieve about 95.5% FMR.

qxh123456789 commented 4 years ago

Hey @XuyangBai , Thank you for your time, I will try this, It sounds considerable, but what is the keypoints number you choose to test? Did your test your registratoin recall? As this version seems a little tricky for me to understand, I will be needing some time to check out the differences between your current architecture and old version and other arguments setting. Will get back to you when I get the test results.

Best, Xianghua Q

XuyangBai commented 4 years ago

Hi @qxh123456789 I mean the FMR is 95.5% over 5000 keypoints, and about 92.6% for 250 keypoints. Sorry I haven't evaluated the registration recall. Yes, the architectures and the arguments are slightly different from the TF version, probably they are not optimal and I just borrow some of them from KPConv-PyTorch or try some values, you may play with them to get better results.