search

GeekAlexis / FastMOT

High-performance multiple object tracking based on YOLO, Deep SORT, and KLT 🚀
MIT License
1.15k stars 253 forks source link

KF coordinates #233

Closed Fschoeller closed 2 years ago

Fschoeller commented 2 years ago

Hi! Thank you for a nice repository. I was wondering why the kalman filter use x1,y1,x2,y2 coordinates and not x,y,s,r like in the SORT trackers. Would x,y,s,r not lead to more consistent predictions as scale and aspect ratio tend to be approximately constant? Have you experimented with this?

Thanks!

GeekAlexis commented 2 years ago

FastMOT uses position coordinates because they are easier to transform when there is camera motion. The constant aspect ratio assumption may help with tracking stability.

rtxbae commented 2 years ago

Hi @GeekAlexis, reading through the state space warping doc as commented in the code, I couldn't understand your implementation (E1, E2, M, w12, w13, etc). Do you refer to any specific publication for that?

Also I would like to implement the state space as x1, x2, s, r, is it remotely possible? Do you have any idea on this? Kalman state space warping literature is almost non-existent regarding this.