Wrong pred_cls_ids Results

[zhujunli1993]

Hi, I'm looking at demo.py codes, and when I ran it to test model's performance on ycbv dataset, I found the pretrained model couldn't segment objects correctly. For example, the GT object class ids should be [1, 4, 6, 19, 20], but the pred_cls_ids was [ 1 6 7 9 13 14 17 18 19 20 21]. I'd like to know how to solve this problem? Thank you!

[ethnhe]

Our segmentation branch is a kind of bottom-up method, which may lead to some false-positive results. There are some ways to filter out these false positives:

• If the objects types in the scene are known as in some manipulation tasks, you can only select the results of target classes.
• If not, ways to reject the false positives include:
  • Increase the threshold of the minimum number of object points here. That may also lead to missing of some heavily occluded objects or small objects that have little visible points.
  • Use the estimated object pose and mesh models to calculate loss with the scene point clouds and filter out objects with a designed threshold.