PRINGLE

PRINGLE ( PeRception INteGrity pipeLinE) is a reproducible research pipeline for ontology-guided semantic validation of vision–language perception in autonomous driving, focused on traffic-light perception under physical manipulation and ambiguity.

PRINGLE is described in:

[article reference]

---

Not included in this repository is the DTLD dataset; you can download it here.

Layout

.
  data/
    dtld/           # YOU provide: DTLD city data
    labels/         # YOU provide: DTLD_Labels_v2.0/v2.0 JSON (Berlin.json, …)
    laser/          # provided: Red/ Green/ IR_modulated/ reference traces
    crops_jpg/      # created by scripts/process_data.py
    synth_jpg/      # optional; created by scripts/process_data.py
  checkpoints/      # place or train *.pt checkpoints here
  output/           # laser ASR + cardinality-consistency results
  scripts/          # Python train / eval / data pipeline

1. Create Environment

conda env create -f environment.yml
conda activate traffic-vlm

2. Prepare data/

Copy (or symlink) inputs into place:

Path	Contents
data/dtld/	DTLD image tree (city folders / *_k0.tiff, …)
data/labels/	Per-city label JSON from DTLD_Labels_v2.0/v2.0/
data/laser/Red, …/Green, …/IR_modulated	Reference laser / IR frames

3. Build crops (+ optional synth attacks)

From this directory:

# Crops only (needed for training / benign eval)
python scripts/process_data.py --crops_only

# Crops + synthetic laser-attack cache
python scripts/process_data.py

# Synth cache only (reuses existing crops / DTLD)
python scripts/process_data.py --synth_only

Defaults:

• --dtld_root -> data/dtld
• --labels_dir -> data/labels
• --laser_root -> data/laser
• --out_root -> data -> writes data/crops_jpg and data/synth_jpg

4. Train

Pretrained models are available upon request.

python scripts/train_models.py \
  --mode baseline \
  --image_dir data/crops_jpg \
  --labels_dir data/labels \
  --save_dir checkpoints/baseline

python scripts/train_models.py \
  --mode card_owned \
  --image_dir data/crops_jpg \
  --labels_dir data/labels \
  --save_dir checkpoints/card_owned

5. Evaluate

From this directory, after checkpoints exist:

bash run_eval.sh

Reads from data/, writes under output/. Default caps are 1000 attack images per combo and 1000 benign images per state. Full eval:

MAX_PER_ATTACK_COMBO=0 MAX_PER_STATE=0 bash run_eval.sh

Optional: BATCH_SIZE, CONDA_ENV.

The results can be visualized with visualize_results.ipynb

Notes

• Override any path with the corresponding CLI flag or env var if your tree differs.
• Shared path constants live in scripts/helper.py (DATA_ROOT, DEFAULT_*).