Vision-based angle measurement using Meta's SAM3 (Segment Anything 3) for press brake applications.

• Text-Prompted Segmentation: Describe what to segment ("bent metal", "sheet metal fold")
• Point Prompts: Click-based segmentation for training/refinement
• Failsafe Logic: Automatic asymmetry detection and correction
• Edge Optimized: Tested on NVIDIA Jetson Orin Nano

• Python 3.10+
• PyTorch 2.7+
• CUDA-capable GPU
• SAM3 (github.com/facebookresearch/sam3)

# Clone
git clone https://github.com/Radar105/sam3-angle.git
cd sam3-angle

# Setup SAM3 environment (if not already done)
git clone https://github.com/facebookresearch/sam3.git ~/sam3
cd ~/sam3 && pip install -e .

# Set SAM3 path
export SAM3_PATH=~/sam3

# Text prompt (primary method)
python main.py --image bend.jpg --prompt "metal" --output result.jpg

# Point prompt (for training)
python main.py --image bend.jpg --point 512,384 --output result.jpg

# With visualization
python main.py --image bend.jpg --prompt "bent metal" -v

from sam3_model import SAM3AngleModel
from image_processor import ImageProcessor
from angle_calculator import AngleCalculator

# Initialize
model = SAM3AngleModel(device='cuda')
processor = ImageProcessor()
calculator = AngleCalculator()

# Measure angle
model.set_image('bend.jpg')
mask = model.get_best_mask(prompt='metal')
processed = processor.process_full_pipeline(mask)
result = calculator.calculate_angle_from_points(
    processed['left_points'],
    processed['right_points']
)

print(f"Angle: {result['final_angle']:.2f}°")

1. SAM3 Segmentation: Text or point prompt identifies the bent material
2. Edge Detection: Canny edges extracted from segmentation mask
3. Point Extraction: Bottom edge points split into left/right groups
4. Line Fitting: RANSAC-style fitting for each side
5. Angle Calculation: Compute angle with failsafe asymmetry correction

Edit config.py:

DEFAULT_PROMPTS = ["bent metal sheet", "V-shaped metal bend"]
FAILSAFE_THRESHOLD = 5.0  # Degrees
DEFAULT_PARAMS = {
    'morph_kernel_size': 5,
    'canny_low_threshold': 50,
    'canny_high_threshold': 150,
    'bottom_margin': 30,
    'percentile': 80,
}

NVIDIA Jetson Orin Nano:

• Model load: ~30s (first run)
• Inference: 5-12s per image
• Memory: ~4GB VRAM

• sam-angle-detection - Original SAM1 version
• obsbot-camera-mcp - Camera control for live capture

MIT