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
378 stars 80 forks source link

Epoch issue #15

Open torchlidar opened 4 years ago

torchlidar commented 4 years ago

Hello.

I am running polarnet. May I ask you how many epochs you have made in the pretrained model?

Thank you.

edwardzhou130 commented 4 years ago

As explained in #3, we didn't set a max epoch number. We trained PolarNet for around 15 epochs and the pretrained model was saved at the 6th epoch.

torchlidar commented 4 years ago

Thank you for your response.

Oh, so if you're going to put it together, you'll just put the train.py on indefinitely and then you'll automatically store the best model.

Thank you!

torchlidar commented 4 years ago

I'm sorry, but I have one more question..!

I would like to ask what mathematical data the Voxel center position calculated.

edwardzhou130 commented 4 years ago

We uniformly cut the space into voxels. The voxel center position is the physical location of each voxel in the space. For the PolarNet, we fixed the space to be distance: 3∼50m, angle: -pi~pi, z:−3∼1.5m, and we set the voxel size as 480,360,32.

thomeou commented 4 years ago

Hi, thank you very much for making the code publicly available. I would like to ask how many epochs did you train the new network (the one with added dropblock and leaklyrelu). I was not able to train your code to match the new pretrained model's performance. Thanks in advance!

edwardzhou130 commented 4 years ago

@thomeou the new network with dropblock requires almost twice the training time than the network without dropblock. In our experiment, it needs around 20 epochs to reach the highest mIoU on the validation split.