AgentI2P: Optimizing Image-to-Point Cloud Registration via Behaviour Cloning and Reinforcement Learning

Round 1

Reviewer 1 Report

This paper proposes an AgentI2P that enables registration across images and point clouds. The authors propose behavioral cloning and reinforcement learning techniques to further improve the policy. Although the algorithm improvement is only for one of the steps, it seems to be useful. There are too many abbreviated proper nouns in this article, so I suggest that the author can construct a separate noun list for easy reading.

Author Response

Point 1: There are too many abbreviated proper nouns in this article, so I suggest that the author can construct a separate noun list for easy reading.

Response 1: Thanks a lot for your suggestion. We add a table in Section 3.5 to describe abbreviated proper nouns in AgentI2P.

Reviewer 2 Report

This paper focused on the image-to-point cloud registration task, where the possible innovations include the newly designed framework for  image-to-point registration via iterative decision process.
In my view, using reinforcement learning is a good idea to handle this problem, and significant performance improvements have been shown from the quantitative comparison table.
Also the effectiveness of the proposed method has been verified by the component table.
Thus, based on the above mentioned aspects, I think this paper is clearly above the acceptance bar of this journal.
Just one MINOR suggestion, the authors shall carefully proof-reading the paper, where I have noticed multiple typos when reading it.

Author Response

Point 1: Just one MINOR suggestion, the authors shall carefully proof-reading the paper, where I have noticed multiple typos when reading it.

Response 1: Thanks a lot for your suggestion. We have carefully proofread the paper and fixed several typos in the manuscript (marked in red).