open-set-segmentation

Open-Set Segmentation Examples

This directory contains examples for open-set segmentation models, which can segment objects based on arbitrary text prompts or visual cues.

Models

SAM3-Image

The latest generation of Segment Anything Model (SAM3) optimized for image-based open-set segmentation.

• Supports both text prompts and geometry (point/box) prompts.
• Advanced hierarchical feature extraction.

Note: For SAM3 with box/point prompts, see the sam3-tracker example.

YOLOEPromptBased

YOLOE with prompt support for flexible object detection and segmentation.

• Visual: Uses a visual prompt (image + bounding box) to find similar objects.
• Textual: Uses text descriptions to segment objects.

Examples

yoloe-prompt-based

# Visual prompt

One image, two boxes prompt

cargo run -r -F cuda-full -F vlm --example open-set-segmentation -- yoloe-prompt-based --model-dtype fp32 --model-device cuda --visual-encoder-dtype fp32 --visual-encoder-device cuda --processor-device cuda --ver 26 --scale s --prompt-image ./assets/bus.jpg -p "xyxy:221.52, 405.8, 344.98, 857.54, person" -p "xyxy:25.519104, 234.08078, 789.1692, 737.5875, bus" --source ../coco2017val/coco/images/val2017 --batch 1

Textual prompt

cargo run -r -F cuda-full -F vlm --example open-set-segmentation -- yoloe-prompt-based \
--model-dtype fp32 --model-device cuda:0 \
--textual-encoder-dtype f16 --textual-encoder-device cuda:0 \
--processor-device cuda:0 \
--ver 26 --scale s \
--source ../coco2017val/coco/images/val2017 --batch 1 \
-p person -p bus 

SAM3-Image

Note: Now sam3image only uses 3 ONNX models (Vision, Text, Decoder with integrated Geometry). No separate geometry encoder needed.

Single text prompt (Use Cuda)

cargo run -F cuda-full -F vlm --example open-set-segmentation -- sam3-image \
--visual-encoder-dtype q4f16 --visual-encoder-device cuda:0 \
--textual-encoder-dtype q4f16 --textual-encoder-device cuda:0 \
--decoder-dtype q4f16 --decoder-device cuda:0 \
--processor-device cuda:0 \
--source ./assets/kids.jpg \
-p shoe

Single text prompt (Use TensorRT)

Note: Use --visual-encoder-dtype f32 for better performance, or --visual-encoder-dtype f16 if VRAM is insufficient. If you encounter conversion failures or memory issues, use --trt_max_workspace_size to increase the TensorRT workspace size (e.g., 8589934592 for 8GB).

cargo run -r -F tensorrt-full -F vlm --example open-set-segmentation -- sam3-image \
--visual-encoder-dtype int8 --visual-encoder-device tensorrt:0 \
--textual-encoder-dtype f16 --textual-encoder-device tensorrt:0 \
--decoder-dtype f16 --decoder-device tensorrt:0 \
--processor-device cuda:0 \
--source ./assets
-p person

Multiple text prompts

cargo run -F cuda-full -F vlm --example open-set-segmentation -- sam3-image \
--visual-encoder-dtype q4f16 --visual-encoder-device cuda:0 \
--textual-encoder-dtype q4f16 --textual-encoder-device cuda:0 \
--decoder-dtype q4f16 --decoder-device cuda:0 \
--processor-device cuda:0 \
--source "assets/kids.jpg" \
-p shoe \
-p "person in blue vest"

Multiple prompts on multiple images

cargo run -F cuda-full -F vlm --example open-set-segmentation -- sam3-image \
--visual-encoder-dtype q4f16 --visual-encoder-device cuda:0 \
--textual-encoder-dtype q4f16 --textual-encoder-device cuda:0 \
--decoder-dtype q4f16 --decoder-device cuda:0 \
--processor-device cuda:0 \
--source "assets/kids.jpg | assets/bus.jpg" -p bus -p cat -p shoe -p cellphone -p person -p "short hair"

Visual prompt (bbox)

cargo run -F cuda-full -F vlm --example open-set-segmentation -- sam3-image \
--visual-encoder-dtype q4f16 --visual-encoder-device cuda:0 \
--textual-encoder-dtype q4f16 --textual-encoder-device cuda:0 \
--decoder-dtype q4f16 --decoder-device cuda:0 \
--processor-device cuda:0 \
--source assets/kids.jpg -p "pos:480,290,110,360"

Visual prompt: multi-boxes prompting(with positive and negative boxes)

cargo run -F cuda-full -F vlm --example open-set-segmentation -- sam3-image \
--visual-encoder-dtype q4f16 --visual-encoder-device cuda:0 \
--textual-encoder-dtype q4f16 --textual-encoder-device cuda:0 \
--decoder-dtype q4f16 --decoder-device cuda:0 \
--processor-device cuda:0 \
--source assets/kids.jpg -p "pos:480,290,110,360;neg:370,280,115,375"

Text + negative box

cargo run -F cuda-full -F vlm --example open-set-segmentation -- sam3-image \
--visual-encoder-dtype q4f16 --visual-encoder-device cuda:0 \
--textual-encoder-dtype q4f16 --textual-encoder-device cuda:0 \
--decoder-dtype q4f16 --decoder-device cuda:0 \
--processor-device cuda:0 \
--source assets/oven.jpg \
-p "handle;neg:40,183,278,21"

Prompt Format

Format	Description	Example
text	Text description	-p "cat"
pos:x,y	Positive point (2 coords)	-p "pos:500,375"
neg:x,y	Negative point (2 coords)	-p "neg:300,400"
pos:x,y,w,h	Positive box (4 coords)	-p "pos:480,290,110,360"
neg:x,y,w,h	Negative box (4 coords)	-p "neg:370,280,115,375"
text;geo;...	Text + geometry	-p "shoe;pos:480,290,110,360"

Parsing Rules

1. First part without pos:/neg: prefix → text prompt
2. Parts with pos:/neg: prefix → geometry (point or box)
3. 2 coords → point, 4 coords → box (xywh)
4. If only geometry (no text), "visual" is auto-set

Examples:

• -p "cat" → text="cat", no geometry
• -p "pos:480,290,110,360" → text="visual" (auto), 1 positive box
• -p "shoe;pos:480,290,110,360" → text="shoe", 1 positive box
• -p "pos:500,375;neg:300,400" → text="visual" (auto), 1 pos point + 1 neg point

Results

model	demo
sam3-image multi-text
sam3-image multi-text
sam3-image visual-pos
sam3-image visual-pos-neg
yoloe-text-prompt
yoloe-visual-prompt