Introducing: a player tracker that's currently #1 on SportsMOT. Training-free, no fine-tuning required
The idea: run a lightweight tracker on every frame, and dispatch SAM 3 only at the moments the tracker is uncertain
Up: Deep-EIoU
Down: SAM-Deep-EIoU (ours)
EXTRA:
A GTA system for the off-screen re-identification is explained in more detail in the blog post and the paper. To get the SOTA score on SportsMOT, SMP and GTA has to be combined (which makes sense, you need a mechanism for off-screen re-id)
Introducing: a player tracker that's currently #1 on SportsMOT. Training-free, no fine-tuning required
The idea: run a lightweight tracker on every frame, and dispatch SAM 3 only at the moments the tracker is uncertain
Up: Deep-EIoU
Down: SAM-Deep-EIoU (ours)
@yacineMTB@suchenzang I mean at some point we need to cut them some slack given that i) the public hated them for adding the guardrails and ii) the admin didn't think they added enough guardrails
Itβs not unprompted, itβs clearly because they believe that this is necessary for them to catch up to the frontier labs, and they value that opportunity above everything else
Not saying it will for sure end up being the best business decision, but the reasoning for why they are doing this is quite straightforward imo
@skalskip92@kr2e I think an accurate touch model + rules (+ the standard stuff that's already solved like court/pitch keypoint detection) would be enough (in e.g soccer and basketball where the ball is visible most of the time) to derive most if not all events that we care about
@jukan05@Alex_Intel_ AMD getting confident in the Helios rollout and feel like they don't need to give away more equity, while Anthropic still wants the same deal OpenAI and Meta got 6 months ago?