search

PRBonn / lidar-bonnetal

Semantic and Instance Segmentation of LiDAR point clouds for autonomous driving
http://semantic-kitti.org
MIT License
959 stars 206 forks source link

The mean and std in config/arch #43

Closed Sabze closed 4 years ago

Sabze commented 4 years ago

Hi, thanks for open sourcing your work! I was wondering how you calculated the mean and std in the config files? When I calculated the mean (for the semanticKITTI sequences 0-10 and 0-21) I got different values than you, especially for the X-coordinate. I got the (range, x, y, z, signal)-means : For the sequences 0-10: (11.6185, -0.1146, 0.4501, -1.0542, 0.2874). For the sequences 0-21: (11.7777, -0.0965, 0.5086, -1.0626, 0.2758). Your mean-values: (12.12, 10.88, 0.23, -1.04, 0.21) It seems reasonable for the mean of the x-value to be around 0, how come it is around 10 for you?

fcyeh commented 4 years ago

Hi, thanks for open sourcing your work! I was wondering how you calculated the mean and std in the config files? When I calculated the mean (for the semanticKITTI sequences 0-10 and 0-21) I got different values than you, especially for the X-coordinate. I got the (range, x, y, z, signal)-means : For the sequences 0-10: (11.6185, -0.1146, 0.4501, -1.0542, 0.2874). For the sequences 0-21: (11.7777, -0.0965, 0.5086, -1.0626, 0.2758). Your mean-values: (12.12, 10.88, 0.23, -1.04, 0.21) It seems reasonable for the mean of the x-value to be around 0, how come it is around 10 for you?

Thanks for the great work. I also wonder why X-coordinate is around 10? I also get around 0. (sequence 00-10 except 08) Here is my mean (range, x, y, z, signal): (11.562, -0.107, 0.502, -1.060, 0.287)

iris0329 commented 4 years ago

I also get x-value around 0. (sequence 00-10 except 08) I also wonder why X-coordinate is around 10?

jbehley commented 4 years ago

Sorry for the delay. I did not personally compute the values, but might still provide some insight and clarification.

I can only say that the exact value is only important when you want to use the pre-trained models. It might be that we somehow used a different subset of the data or just did the computations at 3 am in the morning. It will likely be a mystery that never will be solved.

Use your normalization and the results will be similar, our values are not some "black magic". Don't worry.

Even other projects don't quite know how some normalizations constants were generated, see https://github.com/pytorch/vision/issues/1439. Thus, we are in good company. ;)

tano297 commented 4 years ago

Sorry for the delay on this. Yes, as @jbehley said, don't worry and use your values. It will just converge a bit faster if ours are wrong, it is unlikely the result will change too much. This was calculated, as @jbehley said, at some point 2 years ago using some random script I generated and may totally be wrong, but definitely beats using the unnormalized values, which can go up to 50m, and generate some numerical problems.

tano297 commented 4 years ago

At the time, I was playing around with the code from squeezeseg, and may have used the front part of the scans only, which would explain why there's a significant positive offset in x, but not y or z

jbehley commented 4 years ago

If you still have concerns, please let us know or reopen the issue.

aprilliuwei commented 2 years ago

@Sabze @fcyeh @iris0329 How did you get the two values of img_mean and img_stds, thank you very much.

hahahaff commented 1 year ago

Hi, thanks for open sourcing your work! I was wondering how you calculated the mean and std in the config files? When I calculated the mean (for the semanticKITTI sequences 0-10 and 0-21) I got different values than you, especially for the X-coordinate. I got the (range, x, y, z, signal)-means : For the sequences 0-10: (11.6185, -0.1146, 0.4501, -1.0542, 0.2874). For the sequences 0-21: (11.7777, -0.0965, 0.5086, -1.0626, 0.2758). Your mean-values: (12.12, 10.88, 0.23, -1.04, 0.21) It seems reasonable for the mean of the x-value to be around 0, how come it is around 10 for you?

@Sabze Hello, I would like to know how to calculate your means value in detail. Is it convenient to share your code? Finally, did you replace means value in the config files with the one calculated by yourself？

hahahaff commented 1 year ago

@Sabze @fcyeh @iris0329 How did you get the two values of img_mean and img_stds, thank you very much.

Hello, do you know how to get img_mean and img_stds? If you already know, is it convenient to share with me? Thanks!

iris0329 commented 1 year ago

Hi @hahahaff , just compute the mean and stds by yourself.

@Sabze @fcyeh @iris0329 How did you get the two values of img_mean and img_stds, thank you very much.

Hello, do you know how to get img_mean and img_stds? If you already know, is it convenient to share with me? Thanks!

hahahaff commented 1 year ago

Hi @hahahaff , just compute the mean and stds by yourself. Hello @iris0329 , have you realized calculating the mean and stds by yourself? May I ask you some questions? 1、If I calculate these the mean and stds myself, do I only need to calculate the mean and stds of the projection points on the scale of the image? 2、Because I want to train the network on other dataset，do I have to change the mean and stds? 3、Did you replace the mean and stds in the config files with the one calculated by yourself？ Looking forward to your reply. Thank you very much！