[HN Gopher] AllTracker: Efficient Dense Point Tracking at High R...
___________________________________________________________________
AllTracker: Efficient Dense Point Tracking at High Resolution
Author : lnyan
Score : 45 points
Date : 2025-06-21 17:09 UTC (5 hours ago)
(HTM) web link (alltracker.github.io)
(TXT) w3m dump (alltracker.github.io)
| jauntywundrkind wrote:
| Crazy slick results. Nicely done team!
| upghost wrote:
| > 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
| sheepscreek wrote:
| 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 wrote:
| Apologies, jargon meanings updated.
| markisus wrote:
| 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 wrote:
| 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.
___________________________________________________________________
(page generated 2025-06-21 23:00 UTC)