vobecant / driveandsegment Goto Github PK

hi, authors. Thanks for great work!

Will you release the training code for this work

Hi,

Many thanks for reading my question.

Could you please kindly explain more detailly about how you project the 'ignored' labels from lidar range image to the rgb image. In my opinion, building accurate correspondence between range image and rgb image requires accurate distance measurements, while the distance of the points with 'ignored' labels are invalid.

Thanks again.

Hi,

Thanks for your novel contribution.

Could you clarify the Segment-wise unsupervised labeling part, i.e., section 3.2?

I want to know how you mask out the pixels belonging to the background or another object within the tight rectangle before feeding it to the Vit model. I have tried setting the pixel value to 255 or 0 but feeding such images to the Vit model does not give an informative CLS token for K-means clustering.

Do you perform preprocessing (like PCA) on the CLS token set before performing mini-batch k-means? Also, please comment on the cluster quality obtained from mini-batch k-means. In my case, the cluster quality is not suitable for further processing.

I really appreciate any help you can provide.
Thanks in advance.

Hi,

Many thanks for sharing your excellent work.

The following are some questions about the configurations of the datasets.

1. As stated in Sec. 4.1: 'We train our models on about 7k images from the Waymo'. However, as far as I know, the Waymo training set has around 800 sequences with each sequence containing roughly 200 images. Could you please kindly provide more detailed description about how the images are sampled.
2. Since Waymo dataset contains rgb and range images captured with multiple cameras and lidars(FRONT, TOP, SIDE etc. ). Could you please kindly describe which one you used in the paper.

Many Thanks.