Recognize (SAM3 Experiments)

Short experiments running Meta's SAM3 (Segment Anything Model 3) on images with text prompts, plus a small script to visualize the predicted masks.

ROLE
Explorer
PERIOD
2026
DOMAIN
Computer Vision
STATUS
Published
CODE
GitHub ↗

OVERVIEW

A small set of experiments with Meta's SAM3 (Segment Anything Model 3) for promptable, open-vocabulary segmentation. SAM3 is included as a git submodule; two short Python scripts use it directly, one builds the image model, sets an image, applies a text prompt, and returns masks, boxes, and scores, and the other visualizes the predicted masks. It is a hands-on exploration of the model rather than an original project.

ARRIVED AS

Try Meta's SAM3 for promptable, open-vocabulary segmentation: point it at an image, give it a text prompt, and see what it can segment, as a hands-on way to understand the model rather than a product.

WHAT I BUILT

  1. 01Pulls in Meta's SAM3 as a git submodule and uses its model builder and image processor directly, no reimplementation.
  2. 02A short run script builds the SAM3 image model, sets an image, applies a text prompt (for example 'person'), and reads back masks, boxes, and confidence scores.
  3. 03A companion script visualizes the predicted masks over the image.

WHAT CHANGED

  • A small, working reference for promptable segmentation with SAM3: prompt an image and get masks, boxes, and scores in a few lines.
  • Exploratory by design, this is a learning experiment with Meta's model, not an original system.