search

chaytonmin / Occupancy-MAE

Official implementation of our TIV'23 paper: Occupancy-MAE: Self-supervised Pre-training Large-scale LiDAR Point Clouds with Masked Occupancy Autoencoders
Apache License 2.0
252 stars 18 forks source link

Reconstruction performance on KITTI #9

Closed ZENGXH closed 2 years ago

ZENGXH commented 2 years ago

Hi @chaytonmin, thanks for releasing the code!

I am trying to train the Voxel-MAE model on KITTI dataset, and try to visualize the reconstruction results. Here is a example I got at epoch30: image From top to bottom, it shows the ground truth, the masked input and the model output. It seems that the model didn't reconstruct the input well. I am wondering if this is expected or not? If not, do you have idea which part could be wrong?

Moreover, may I ask which dataset are you using in Figure 3.

chaytonmin commented 2 years ago

Hi @chaytonmin, thanks for releasing the code!

I am trying to train the Voxel-MAE model on KITTI dataset, and try to visualize the reconstruction results. Here is a example I got at epoch30: image From top to bottom, it shows the ground truth, the masked input and the model output. It seems that the model didn't reconstruct the input well. I am wondering if this is expected or not? If not, do you have idea which part could be wrong?

Moreover, may I ask which dataset are you using in Figure 3.

We use KITTI in Figure 3.

How do you visualize the reconstruction results? What is the masking ratio?

ZENGXH commented 2 years ago

Thanks for your reply! I found the issue: it's because I only plot the top-5000 points in the reconstruction image. If I increase the number it looks better.

huixiancheng commented 11 months ago

Hi, xiaohui. @ZENGXH Would you like to share your vis code ? If it still exists. :joy_cat: I'm confused about getting points from dense's features.