search

edwardzhou130 / PolarSeg

Implementation for PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation (CVPR 2020)
BSD 3-Clause "New" or "Revised" License
373 stars 80 forks source link

train epoch #27

Open Zhang-jiemmd opened 3 years ago

Zhang-jiemmd commented 3 years ago

I want to know how many epochs you trained on the kitti dataset I have trained 15 epochs,and best val miou is 47.082

edwardzhou130 commented 3 years ago

As mentioned in #15, training with dropblock needs around 20 epochs to reach the highest mIoU.

edwardzhou130 commented 3 years ago

But 47.082 at the 15th epoch is much lower than what I got. Are you using the default training settings in the code?

Zhang-jiemmd commented 3 years ago

I set grid_size=[320, 320, 16], model='traditional'. Other parameters are the same as the default values. Does the difference in such parameters have any effect?

edwardzhou130 commented 3 years ago

Smaller grid_size generally has a lower mIoU, but it won't affect too much.

traditional model trains the same network in cartesian coordinates rather than in polar coordinates. Our experiment shows that its mIoU performance is lower than PolarNet.