fabiopoiesi / gedi Goto Github PK

I ported the training code from the dip algorithm. Ask gedi if he suffers from a single gpu and false limits during training.

where is your code?

When running demo.py, it takes 8-10 seconds for gedi.compute() to execute. (Run on GTX 1660 Ti)

The paper mentions that Gedi runs in 1.454ms, but does demo.py achieve a similar speed with its settings?
If the settings differ, what is the approximate execution speed for demo.py?

Hello, what about the rest of the code?

Thank you for your excellent work. Do you have any plans to upload the training code or the checkpoint trained on the KITTI dataset?

Hi, I really appreciate your work on point cloud registration . Is there any estimated date for the code release? Thank you very much.

Hi, thanks for your great work!

I am very interested in your work, and trying to re-implement it. Could you share the settings of hyper-parameters in each layer? I wonder how you set the radius size and the number of sampling points in pointnet2_utils.QueryAndGroup.

Looking forward to your reply. Thank you very much.

Best,
Sheng

Is there a lack of steps to train the network?

I am planning to use the repository to register my own data. Can you please provide steps to train the network?

Hey there!
I am struggling with using gedi on small point clouds (100 - 500 points). Default hyperparameters and 3dmatch checkpoint don't let me get a good registration. Do you have any ideas? How can I change hps to work with smaller point clouds?

Best regards and thanks for the work!

Hi,

you say in your work the following:

We use the point cloud pairs whose overlap is τo ≥ 30%, resulting in about 16.6K pairs for training and about 1.66K pairs for testing [8], [10], [16]. We use the ground-truth transformations provided by Gojcic et al. [16]

Could you please tell me how do you get the 1.66K pairs for testing?
There are 1632 registrations to complete in the official 3DMatch benchmark.

Do you compute the overlap for every possible pair of point clouds in a fragment and, if the overlap is > 30%, try to register those examples? As I presume Gojicic et al. did in the PREDATOR paper here (and not in [16] Perfect match paper)?

Gojcic et al. [16] do not have any data either on ther repository or on their download links so I am a bit confused.

Thank you in advance and kind regards,
David

Why is the registration effect of this model not good on my own point cloud data? Do I need to modify some parameters and hope to get some advice

Hello, could you please provide the data of Figure 4 in the paper? Thank you!

hello，cloud you tell me what's the voxel_size of kitti in preprocess_dataset,please.
or provide the kitti's preprocess like preprocess_3dmatch_lrf_train.py