Vision QA bot on a budget: Segment Anything + supervision, both free on the catalog

I've been building a visual quality-assurance tool for a small electronics manufacturer — they want to flag obvious PCB defects before boards go into final assembly. Commercial CV APIs were quoted at $0.004 per image, which adds up fast at 5,000 boards a day. I went looking for a free path and found it in the catalog.

The two free listings

• segment-anything-model — Meta's SAM; zero-shot segmentation, no class-specific training required
• supervision-vision-toolkit — Roboflow's annotation and detection utilities; great for drawing detections and computing IoU metrics

Both free to install, both security-scanned with no issues. The listing security tabs show the scan report clearly — for ML model files that's especially useful since the scanner checks for pickle-based exploits and hidden execution in the model format layer.

The setup

import cv2
import numpy as np
import supervision as sv
from segment_anything import sam_model_registry, SamAutomaticMaskGenerator

# Load SAM (vit_b checkpoint, ~375 MB)
sam = sam_model_registry["vit_b"](checkpoint="sam_vit_b.pth")
mask_generator = SamAutomaticMaskGenerator(
    model=sam,
    points_per_side=16,
    pred_iou_thresh=0.88,
    stability_score_thresh=0.95
)

def inspect_board(image_path):
    img = cv2.imread(image_path)
    img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

    masks = mask_generator.generate(img_rgb)

    # Flag segments with unusually small or irregular area as potential defects
    defects = [m for m in masks if m["area"] < 200 or m["stability_score"] < 0.7]

    # Annotate with supervision
    detections = sv.Detections.from_sam(masks)
    annotator = sv.MaskAnnotator()
    annotated = annotator.annotate(scene=img_rgb.copy(), detections=detections)

    return annotated, len(defects)

result_img, defect_count = inspect_board("board_0042.jpg")
print(f"Potential defects: {defect_count}")

Numbers after two weeks of testing

• Defect detection rate: ~73% on a labelled test set of 800 images (false negative rate still a bit high for solo use, but good as a first-pass filter)
• False positive rate: ~18% — acceptable for "flag for human review"
• Throughput: ~1.1 images/second on a single T4 GPU (rented at $0.35/hr only during batch runs)
• Effective cost per image: < $0.0001

Compared to the commercial API quote, we're saving ~$4,900/month at current volume.

The catalog's vision category has more listings building on this foundation — if you're doing any CV work, it's worth a browse before reaching for a paid API.