Purpose

Segment SAM3 finds regions that match a concept such as person. It produces masks that follow the subject, plus detection rows and a view drawn over the image. Spatial point or box prompts can further guide the result.

Use

Place Segment SAM3 under Analyze in the Image view and connect footage to image_in. Start with one word in Concepts and inspect preview. Adjust Confidence until the intended subjects remain.

Select Mask in Field and use field downstream as a mask. Enter comma-separated concepts, such as person, chair, when you need several kinds of subject.

Inputs and outputs

• image_in is the source picture.
• prompts optionally supplies point or box prompt rows.
• field publishes the selected mask, edge distance, identity or confidence view.
• preview draws the selected visualization over or in place of the source.
• detections contains per-instance rows.
• count and summary give the instance count and combined statistics.

Controls

• Concepts names the subjects to find.
• Confidence sets the acceptance threshold. Max Instances limits the retained results.
• Preview selects Composite, Outlines, Masks or another visualization.
• Field selects what the field image encodes. Line Width controls drawn outline thickness.
• Image Fit selects Squash or Letterbox when preparing the picture for the model.

Notes

Check the mask on the actual footage before using it to isolate a subject. The reported track IDs are best-effort matches between frames and can change. Use Track Anything to follow chosen points through a sequence.

Ports

Inputs

• image_in image required
  RGBA
• prompts image
  64 x 8

Outputs

• field image
  RGBA
• preview image
  RGBA
• detections image
  64 x 19
• count image
  1
• summary image
  1 x 14

Parameters

Segment

Concepts (Prompt) Str op('segment_sam').par.Prompt

Default:

person

Confidence (Threshold) Float op('segment_sam').par.Threshold

Default:

0.15

Range:

0.01 to 0.99

Max Instances (Maxinstances) Int op('segment_sam').par.Maxinstances

Default:

16

Range:

1 to 64

Preview (Previewstyle) Menu op('segment_sam').par.Previewstyle

Default:

composite

Options:

composite, outlines, masks, source, instance, concept, confidence

Field (Fieldstyle) Menu op('segment_sam').par.Fieldstyle

Default:

mask

Options:

mask, edge_distance, instance_id, concept_id, confidence, query_id, coverage

Line Width (Previewthickness) Int op('segment_sam').par.Previewthickness

Default:

3

Range:

1 to 16

Image Fit (Fit) Menu op('segment_sam').par.Fit

Default:

squash

Options:

squash, letterbox