Questions on Size Offsets and Objectness

[lyp0413]

Hello!May I have some questions below:

1.Why it predict the size offset rather than predicting the size directly? 2.My objectness are all very large on my own dataset (nearly 1),what are the possible reasons?

Looking forward to your help!

Thank U!

[NUAAXQ]

For the first question, votenet use the same residual (offset) regression strategy in Frustum PointNets. As reported in Frustum PointNets, they follow the idea of Faster R-CNN and 3D-Deepbox. The original explanation in 3D-Deepbox is here: In all, it is believed that it is easier to get good results by using the residual regression strategy than direct prediction. As for the second question, it is hard to give the reason without seeing the dataset. Hope the above can help you.