mit-han-lab / pvcnn

[NeurIPS 2019, Spotlight] Point-Voxel CNN for Efficient 3D Deep Learning

https://pvcnn.mit.edu/

MIT License

644 stars 130 forks source link

visualization and features #61

Closed sunghyun-nam closed 2 years ago

sunghyun-nam commented 2 years ago

Hello. I'm studying for 3d semantic segmentation and I'm new in this field. So I don't know what shoud I do for visualization using your model. I also watched your advice and reference code related to visualization. But I couldn't understand. Among the answers you gave, you were like "The visualization is based on the features from the last PVConv. You may color each point based on the relative L2-norm of its features from either point-based or voxel-based branches." I know what you mean.. But I don't know how to use features. Thanks in advance.

54

zhijian-liu commented 2 years ago

Could you be more specific about your question? What do you want to visualize? Which step are you currently at? Thanks!

sunghyun-nam commented 2 years ago

Could you be more specific about your question? What do you want to visualize? Which step are you currently at? Thanks!

I'd like to get the results in figure 6

54

I just do the process in README.md. There're just how to train(test, evaluate) model. I want to visualize the results(features) from the model.

zhijian-liu commented 2 years ago

You could follow https://github.com/mit-han-lab/spvnas/blob/master/tutorial.ipynb to visualize the point cloud. For feature visualization, you need to generate the color based on the L2 feature norm.