Hi SAM 3 team,
Thank you for releasing SAM 3, the model checkpoints, code, and the SA-Co evaluation benchmarks.
I would like to ask about the release plan for the training datasets used by SAM 3, especially:
SA-Co/HQ
SA-Co/SYN
SA-Co/VIDEO
Could you please clarify:
Is there currently a plan to publicly release these training datasets?
If yes, is there an estimated release timeline?
How much of each dataset is expected to be released?
For example, will the full datasets be available, or only a subset?
Will the release include the original images/videos, annotations only, or image/video identifiers and download scripts?
Will the released annotations include the noun phrases, masks, hard negatives, and other metadata used during SAM 3 training?
Are there licensing or redistribution constraints that may prevent the full SA-Co training data from being released?
The paper provides useful statistics about the scale of SA-Co, but for researchers interested in reproducing or extending SAM 3, it would be very helpful to understand what portion of the training data will eventually become publicly accessible.
Even an approximate roadmap or clarification on which components are expected to remain private would be very useful.
Thank you!
Hi SAM 3 team,
Thank you for releasing SAM 3, the model checkpoints, code, and the SA-Co evaluation benchmarks.
I would like to ask about the release plan for the training datasets used by SAM 3, especially:
SA-Co/HQ
SA-Co/SYN
SA-Co/VIDEO
Could you please clarify:
Is there currently a plan to publicly release these training datasets?
If yes, is there an estimated release timeline?
How much of each dataset is expected to be released?
For example, will the full datasets be available, or only a subset?
Will the release include the original images/videos, annotations only, or image/video identifiers and download scripts?
Will the released annotations include the noun phrases, masks, hard negatives, and other metadata used during SAM 3 training?
Are there licensing or redistribution constraints that may prevent the full SA-Co training data from being released?
The paper provides useful statistics about the scale of SA-Co, but for researchers interested in reproducing or extending SAM 3, it would be very helpful to understand what portion of the training data will eventually become publicly accessible.
Even an approximate roadmap or clarification on which components are expected to remain private would be very useful.
Thank you!