Inconsistent code and paper description

[skylerlxy]

Formula 2 in the paper sums up the clip features of each point and performs l2 regularization, but in the code, the clip features that undergo l2 regularization are summed up. Will this lead to the constant addition of clip features and an increase in the clip feature modulus of frequently accessed points?

After obtaining the summation feature, there is no subsequent regularization step.

[PhucNDA]

@skylerlxy
Thank you for your interest in our work. I believe you're referring to this line: refine_feat. In the first stage, we use grounding_2d to generate the bounding box, mask, and feature. Then, we apply refine_feature to recompute the features before performing the final point cloud feature normalization (via summation). The resulting processed features are then used for open-vocabulary querying.
I hope this help you clarify your cocern.

Best,
PhucNDA.