Text-Prompted Video Object Tracking
Grounded-SAMType what to track — Grounding DINO + SAM2 mask it through the whole clip, not SAM3
Upload a short video and describe the object to track
Clips are trimmed to 6s and processed at 4fps, downscaled — CPU inference on a free hosted Space is slow. Expect roughly 15-40s depending on clip length.
What this doesn't do: this is not Meta's SAM3 — SAM3's checkpoints are currently gated behind a Meta access request, so this uses SAM2 + Grounding DINO ("Grounded-SAM") instead, both freely available. The output is a sampled-frame preview (reduced fps/resolution), not a full-resolution exported video file, and Grounding DINO only looks at the first frame. It also has no reliable way to say "nothing here matches" — a wrong description can still return a plausible-looking (but wrong) tracked region instead of an error, so always check the result visually.