Hello author, First of all, please forgive my nagging. I hope i won't make you feel bored.

• IdM_1's rotation is identity, its translation is - rotation_element.rot_origin, calculated in rot_element.rot_origin = vec_first_t;, namely -P_j. IdM_2 is R_z. In paper, p_zk = Rz * p_k, q_k' = transform * q_k; q_zk = Rz * q_k'. There is no IdM_1 variable in the paper. Q1: What's the meaning of IdM_1?

IdM_1.block<3, 1>(0, 3) = -1.0 * origin;
IdM_2.block<3, 3>(0, 0) = twoVectorsAlign(axis, z_axis);


Eigen::Matrix4f local_tm_t = IdM_2*IdM_1;
Eigen::Matrix4f local_tm_s = local_tm_t*transform.matrix();


pcl::transformPointCloud(*key_source_, *source_local, local_tm_s);
pcl::transformPointCloud(*key_target_, *target_local, local_tm_t);

• Q2: I do not totally understand the func vl_fast_atan2_f(), why not directly use atan2() to calculate the coordinate azimuth, is there something we need to consider?
• Q3: The relationship of R(theta, epsilon_PQ_ij), R(theta, Z) and Rz is R(theta, epsilon_PQ_ij) = R(theta, Z) * Rz ?
• Q4: Related to Q1, the final translation t's calculation seems is different from eq(18) of paper, also confused this.

Eigen::Matrix4f IdM_1 = Eigen::Matrix4f::Identity();
Eigen::Matrix4f IdM_2 = Eigen::Matrix4f::Identity();
Eigen::Matrix4f IdM_3 = Eigen::Matrix4f::Identity();
Eigen::Matrix3f rot = Eigen::AngleAxisf(best_angle, two_point_rot_ele.rot_axis).toRotationMatrix();
IdM_1.block<3, 1>(0, 3) = -1.0*two_point_rot_ele.rot_origin;
IdM_2.block<3, 3>(0, 0) = rot;
IdM_3.block<3, 1>(0, 3) = two_point_rot_ele.rot_origin;
Eigen::Matrix4f gr_tran_mat = IdM_3*IdM_2*IdM_1*two_point_tran_mat;

• BTW, in code best_count_ = der_in_tcfs;, in paper pseudo code line20 of "Algorithm 1" is best_count_ = der_in_tcfs + 2; I don't understand why plus 2?

Thanks for author's patience and time in advance sincerely!

Hello author, Firstly, thank you for sharing your excellent work to the community to let us study. There are many symbols in the paper, thank you for your meticulous and rigorous.

• Q1: When calculating calEdgeReliabilityInRCFS(), we use V_Pk and V_Qk in code and the same with paper eq(11). But in the pseudo code line5 of "Algorithm 1: Maximum consensus set based on edge reliability", using transformed V_Qk, is it a typo in paper?

	for (int j = 0; j < correspondences.size(); ++j)5 months ago • WeiPengCheng [outlier removal for point cloud registration]
	{
		auto& t_p = (*key_target_)[correspondences[j].index_match];
		auto& s_p = (*key_source_)[correspondences[j].index_query];


		diff_to_t[j] = t_p.getVector3fMap() - first_corr_t.getVector3fMap();
		diff_to_s[j] = s_p.getVector3fMap() - first_corr_s.getVector3fMap();
	}

• Q2: in twoPairPointsAlign(), it seems the first pair is point j, the second pair is point i. epsilon_p_ij = pj - pi; When in calEdgeReliabilityInRCFS() we calculate epsilon_p_ik = pk - pi, epsilon_q_ik = qk - qi. Maybe use diff_to_t[j] = t_p.getVector3fMap() - second_corr_t.getVector3fMap(); and diff_to_s[j] = s_p.getVector3fMap() - second_corr_s.getVector3fMap(); more consitent with paper, although its work is the same with minus the first pair.

Eigen::Vector3f vec_sour = (vec_first_s - vec_second_s).normalized();
Eigen::Vector3f vec_tart = (vec_first_t - vec_second_t).normalized();

Hi Pengcheng Wei, Thanks for your paper. I was wondering if your method could be adapted for a 3DOF version that only requires translation estimation? If not, can you recommend any of the other techniques you have tried that might be more accurate / efficient under the 3DOF case? ( I could not find your email address so I made an issue, feel free to close )

Thanks for your sharing, I got uncorrect result when I test it with the following command. ./GrorReg "arch/s1.ply" "arch/s5.ply" 0.1 800

the two clouds are not aligned as in your paper.

could be related to parameter settings?