The Plan

• Pick a vehicle. Any camera works — the paper used an iPhone.
• Shoot in: full sun, overcast, dusk, parking lot, street parking.
• Vary distance (10–60 ft) and angle (straight-on, 15°, 30°).
• Drop the .jpg files into /home/ubuntu/spar/data/raw/

Option A — Web GUI (recommended for teams):

cd /home/ubuntu/spar
source venv/bin/activate
python web_labeler/labeler_server.py --images data/raw/ --labels data/labels/ --port 5050

Open http://YOUR_SERVER_IP:5050 in any browser. Everyone enters their name, clicks corners simultaneously, images auto-assign so no overlap. Full controls on the Web Labeler page.

Option B — Desktop tool (single user):

python label_corners.py -i data/raw/

Click the 4 corners in order: Top-Left → Top-Right → Bottom-Right → Bottom-Left. ~3 seconds per image = 300 images in ~15 minutes. Output: JSON files in data/labels/.

Verify labels:

python check_labels.py -i data/raw/ -l data/labels/

python train_patch.py \
  --mode disruption \
  --epochs 100 \
  --batch-size 64 \
  --lr 0.01 \
  --tv-weight 0.01

What happens:

• Loads your 300 photos + corner labels; initializes a random patch (Xavier init).
• Per image: warps the patch via kornia homography → composites as rim → runs through YOLO + OCR → computes loss → backpropagates to the patch.
• Three loss terms fight each other:
  • Detection loss — makes YOLO unsure / miss the plate
  • OCR loss — makes the text reader output gibberish
  • TV loss — keeps the patch smooth and printable
• Saves checkpoints every 20 epochs.

Outputs in patches/:

python train_patch.py \
  --mode impersonation \
  --target "ABC1234" \
  --epochs 100 \
  --batch-size 64

Makes the ALPR read your plate as a different plate text instead of failing to read it.

• Take patches/patch_disruption_epoch100_printable.png to a professional print shop (FedEx Office, Staples, local shop).
• Paper: heavyweight matte poster · Size: 8.4″ × 16.8″ (already 300 DPI = 2520×5040 px) · Cost: ~$50
• Cut out the center so the plate face is visible — only the border/rim remains.

• Place the printed rim between the license plate and the vehicle.
• The rim extends outward from the plate edges.
• No part of the rim covers any letter, number, county name, or state name.
• Secure it so it doesn't flap in the wind.

python evaluate.py \
  --patch patches/patch_disruption_epoch100.pt \
  --images data/eval/ \
  --labels data/labels/ \
  --output results/

What to put in data/eval/:

• Fresh photos of the car with the rim installed — same conditions as training.
• Also the same shots WITHOUT the rim (control).

What you get back (results/):

To render a standalone patch later:

python render_patch.py --patch patches/patch_disruption.pt --output patches/new_print.png