SAM Video Annotation AI

Segmentation labels are the slowest labels to draw by hand, and video multiplies the work by every frame you keep. Roboflow puts Meta's Segment Anything models into the labeling flow, so you prompt and review instead of tracing outlines one point at a time.

Turn video into frames:

• Upload an .mp4 or .mov to your project and extract frames at the sampling rate you choose, like one frame per second or every fifth frame.
• Pick a lower rate to cut near-identical frames and annotation work, or a higher rate when objects move fast.

Label with SAM 3:

• Run Auto Label with SAM 3 (opens in a new tab) to label objects from text prompts, and test prompts and thresholds on sample frames before you run the job. Test results are free.
• Pair a foundation model with SAM 3, like Gemini 4 Argon + SAM 3 or GPT-6 Astra + SAM 3, to label from class descriptions and get instance segmentation masks.
• Starting small? See how to grow small datasets with SAM 3 using auto-labeling, mask validation, and segmentation-aware augmentation.

Refine and track:

• Fix any mask with Smart Select (opens in a new tab), which runs SAM in your browser, and use Repeat Previous to carry the last frame's annotations forward for a quick adjust.
• Track objects through a video in a Workflow with the SAM 3 Video Tracker (opens in a new tab) block, which segments and tracks objects from text prompts, or the SAM 2 Video Tracker, which follows boxes from an upstream detector. Both return masks with tracker IDs and run on a GPU.
• Go deeper on video segmentation with SAM 2.

Start building today. Your first segmentation dataset is one video away.

‍

Frequently asked questions

What is SAM video annotation?

SAM video annotation uses Meta's Segment Anything models to label objects in video frames with pixel-accurate masks, from a text prompt or a click, instead of drawing every outline by hand.

Do I have to label every frame?

No. Sample frames at a rate that captures how your objects move, auto-label the batch with SAM 3, and review the results, fixing only the masks that need it.

Should I use SAM 3 or SAM 2?

Use SAM 3 when you want to segment and track objects from text prompts, like "forklift" or "pallet". Use SAM 2 when you already have a detector and want to turn its boxes into tracked masks.

What can I train on the results?

Instance segmentation models like RF-DETR Segmentation, or object detection models. In an object detection project, masks are saved as boxes for training.

‍