search

nubot-nudt / MDGAT-matcher

[RAL] Keypoint Matching for Point Cloud Registration Using Multiplex Dynamic Graph Attention Networks
MIT License
38 stars 14 forks source link

Info regarding number of keypoints for training #8

Closed vishalsharbidar closed 2 years ago

vishalsharbidar commented 2 years ago

Hello, Thanks for the work and great paper.

While training, I noticed that the number of keypoints for img1 and img2 is constant. The loss function is also built on this assumption. Can you please let me know what may happen if I use a different number of keypoints for img1 and img2? Would it have an adverse effect on the training? For example: If I use 500 keypoints for Img1 and 1000 keypoints for Img2.

Thank you.

chenghao-shi commented 2 years ago

Hi，

We keep the number of keypoints to be constant for the convenience of batch training. It's no need for the number for img1 to be the same as the number for img2. The training is feature-level, more keypoints mean more samples, so I don't think it will have an adverse effect. But after increasing the number of keypoints, the memory usage of the attention increases significantly, which is probably the only thing you need to be aware of. 

施成浩 @.***

 

------------------ 原始邮件 ------------------ 发件人: "chenghao-shi/MDGAT-matcher" @.>; 发送时间: 2022年9月4日(星期天) 晚上8:06 @.>; @.***>; 主题: [chenghao-shi/MDGAT-matcher] Info regarding number of keypoints for training (Issue #8)

Hello, Thanks for the work and great paper.

While training, I noticed that the number of keypoints for img1 and img2 is constant. The loss function is also built on this assumption. Can you please let me know what may happen if I use a different number of keypoints for img1 and img2? Would it have an adverse effect on the training? For example: If I use 500 keypoints for Img1 and 1000 keypoints for Img2.

Thank you.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>

vishalsharbidar commented 2 years ago

Thanks for the input, It was helpful.

vishalsharbidar commented 2 years ago

So, when I have a different number of keypoints for image pairs, would it be good to use batch size = 1 for training?

chenghao-shi commented 2 years ago

I think using a larger batch size will give better results.

施成浩 @.***

 

------------------ 原始邮件 ------------------ 发件人: "chenghao-shi/MDGAT-matcher" @.>; 发送时间: 2022年9月5日(星期一) 晚上9:28 @.>; @.**@.>; 主题: Re: [chenghao-shi/MDGAT-matcher] Info regarding number of keypoints for training (Issue #8)

So, when I have different number of keypoints for image pairs, would it be good to use batch size = 1 for training?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

vishalsharbidar commented 2 years ago

Okay, Thank you