AllTracker: Efficient Dense Point Tracking at High Resolution
7 comments
·June 21, 2025upghost
> The utility of optical flow (i.e., the instantaneous velocity of pixels [16]) toward this goal has long been obvious, yet it has remained challenging to upgrade flows into long-range tracks.
This sentence from the paper makes me feel a little bad that I don't understand why this goal is obvious. I am not tracking why we are tracking pixels.
Is this basically a competing technology with YOLO[1] or SAM[2]?
[1]: https://en.m.wikipedia.org/wiki/You_Only_Look_Once
[2]: https://ai.meta.com/sam2/
Edit: added annotations, should've done that initially
markisus
Back in my earlier days working on autonomous vehicles, I dreamed of something like this.
The issue with bounding boxes is missed detections, occlusions, and impoverished geometrical information. But if you have a hundred points being stably tracked on an object, it's now much easier to keep tracking it through partial occlusions, figure out its 3D geometry and kinematics, and even re-identify it coming in and out of occlusion.
daemonologist
No, this performs the same task as CoTracker or TAPIR, but intended for running at a higher resolution. Point tracking is useful both for keeping track of the position of a target and for "inside-out" positioning of the camera.
YOLO is mostly concerned with detecting objects of certain classes in a single image, and SAM is concerned with essentially classifying pixels as belonging to an object or not.
sheepscreek
I’m not remotely familiar with either YOLO or SAM, but want to add my own question here. Does the utility of this invention have something to do with the tracking of subjects, like auto-focus for cameras and robotics (to keep the subject in view)?
upghost
Apologies, jargon meanings updated.
jauntywundrkind
Crazy slick results. Nicely done team!
Object segmentation and tracking is such a natural and 'automatic' part of our visual perception that it's difficult to intuit how challenging it is to do with software.