Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Note for training #6

Closed
Sciroccogti opened this issue Aug 26, 2019 · 4 comments
Closed

Note for training #6

Sciroccogti opened this issue Aug 26, 2019 · 4 comments

Comments

@Sciroccogti
Copy link
Member

Sciroccogti commented Aug 26, 2019

Useful issues on YOLO6D:

  1. Huge translation error but reasonable pixel error and ACC measurement?:
    This issue is opened by F2Wang who is the author of the dataset making tool which is being used by us, so it could be pretty useful.
  2. Unofficial F.A.Q.
@Sciroccogti
Copy link
Member Author

Sciroccogti commented Aug 26, 2019

How to get a good result

Parametres

There are several parameters you need to modify when you are about to train. However, usually these parameters won't change too much as long as you are using the same camera and the same dataset making tool.

  1. Camera Intrinsics
    The dataset making tool which we are using can output the intrinsics. Strictly speaking, it is important for you to change the intrinsics everytime you start to train a new dataset. However since you are using exactly the same camera, the intrinsics won't vary too much to affect the result.

  2. Diametre
    Diametre is the maximum distance among the corners of the 3D bounding box. Obviously this parametre changes a lot between different objects. While you are using the dataset making tool, you should have noticed that the tool provides a way to generate the diametre of the current object. However, the valid.py itself can calculate the diametre, so commonly there's no need to modify this parameter manually.

  3. batch_size
    Batch is a common parametre in CNN. As we all know (if you don't, try to train on your own laptop and you will find out🙃), batch_size is limited by the memory of your GPU. Try to cut down on the batch_size when facing error CUDA out of memory or something else. Meanwhile, it is said that the bigger the batch_size is, the better the result will be. So make sure that you have a appropriate batch_size. Notice: we are using 8 on Aliyun servers.

  4. anchor
    See also. I have never changed this one yet.

@Sciroccogti
Copy link
Member Author

How to obtain a better result

  1. More pictures
    This way is always useful.

  2. Better pictures
    Make sure that enough aruco markers can be seen in each picture.

  3. Carefuller .ply modifying
    See also

@Sciroccogti
Copy link
Member Author

How to set a GPU server

I chose aliyun.
The system disc should be at least 20GB, and the data disc should be at least 40GB.
The reference price is ¥12 per hour.
REMEMBER TO TURN OFF THE SERVER WHEN NOT USING THEM!

@Sciroccogti
Copy link
Member Author

Sciroccogti commented Sep 2, 2019

Moved to wiki

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant