a question about ATE RMSE results on ScanNet

[wangm-buaa]

Hi, thanks for your excellent work!
I have a question after evaluating scene0000_00 with the code you provided. There are two types results pose_x.png and pose_r_x.png generated by following codes.

Co-SLAM/coslam.py

Lines 675 to 676 in 2329d09

	pose_evaluation(self.pose_gt, self.est_c2w_data, 1, os.path.join(self.config['data']['output'], self.config['data']['exp_name']), i)
	pose_evaluation(self.pose_gt, pose_relative, 1, os.path.join(self.config['data']['output'], self.config['data']['exp_name']), i, img='pose_r', name='output_relative.txt')

It seems that using pose_relative can achieve better ATE RMSE result, while the other result is more similar as the one on paper.
I wonder which result is the one you provided on paper and why you didn't use the other one.
Thank you!

[HengyiWang]

Hi @wangm-buaa, thank you for your interest in our work. We use pose_relative in the paper. The difference between those two is that in pose_relative, we save the relative poses of the non-keyframe with respect to the previous keyframe while in est_c2w_data , the non-keyframe poses are saved with respect to the first frame. Keyframe poses will be updated after performing BA, then all poses in pose_relative would change. Most of the time, the results using pose_relative are better meaning that BA is working very well. Thus, I keep both results in the code :)

Please note that in ScanNet, we run each scene 5 times and report the average results. Also, I did some refactoring of the code in the released version, so you might get slightly better tracking results.

[wangm-buaa]

Thanks for your reply and help!

[ArlenCHEN]

Following your discussions, is it for Co-SLAM to have better results if we use pose_relative instead of est_c2w_data to perform the ray tracing in tracking_render and global_BA?

[HengyiWang]

Hi @ArlenCHEN, we only use the keyframe rays and the corresponding poses for global BA.

[ArlenCHEN]

Thanks @HengyiWang.

Another question: How do you generate the online reconstruction video? Did you release the code for it?