search

Haiyang-W / CAGroup3D

[NeurIPS2022] This is the official code of "CAGroup3D: Class-Aware Grouping for 3D Object Detection on Point Clouds".
89 stars 9 forks source link

About processed data of SUN-RGBD #11

Closed Yukichiii closed 1 year ago

Yukichiii commented 1 year ago

Hi, I noticed that in cfgs/dataset_configs/sunrgbd_dataset.yaml, you sampled 100000 points. But the data in your processed data only have 50000 points. And in FCAF3D, they mentioned that "The only difference is that we do not sample 50,000 points from each point cloud in SUN RGB-D, using all points." So, I wonder if this affects your experimental results? Which version of data can reproduce you result?

Haiyang-W commented 1 year ago

Sorry to reply then, I am rushing some ddl. It's a very good question. We reproduce the sunrgbd without 50000 points constraint. Apologize. We will update the data as soon as possible.

Haiyang,

Haiyang-W commented 1 year ago

That would be a boost of about 0.5. We will update as soon as possible.

Haiyang-W commented 1 year ago

The new version of SUN RGBD data is updated. Remove the indoor sampling step, and you can easily achieve the SOTA performance of SUN RGBD based on the new data. The training log and model weight will be updated in one day.