Umeyama alignment and scaling for computing ATE-5

tinghuiz / SfMLearner

An unsupervised learning framework for depth and ego-motion estimation from monocular videos

MIT License

1.97k stars 557 forks source link

Umeyama alignment and scaling for computing ATE-5 #125

Open hemangchawla opened 4 years ago

hemangchawla commented 4 years ago

Hello,

Why is the standard Umeyama algorithm not used for alignment and scaling of 5-frame trajectories when computing the ATE-5?

adizhol commented 4 years ago

I would also like to understand that. The Umeyama algorith uses the SVD of the covariance matrix to calculate the alignment between two trajectories. http://rpg.ifi.uzh.ch/docs/IROS18_Zhang.pdf

edit One possible reason is that finding alignment of the entire trajectory estimation will give you a different error (kind of an "average"), but if the alignment is done on the first frame pose, the errors are accumulated over the entire trajectory. Correct me if I'm wrong.