Prepare ZOD¶

Install Fusion first, with [visin] for the download (python -m pip install 'visin-fusion[visin]'). Run from a source checkout, since the example configs live there.

Download and prepare¶

Copy this block into your terminal:

Prepare ZOD

export VISIN_DATASET_URL=https://dataset-api.visin.eu
export VISIN_DATA_DIR=./data
ZOD_DATA_DIR="$(visin download zod)"
export ZOD_DATA_DIR
if [ ! -f "$ZOD_DATA_DIR/dataset.json" ]; then
  visin-fusion dataset manifest \
    --root "$ZOD_DATA_DIR" \
    --config configs/datasets/zod.json
fi

The download prints the extracted folder; progress goes to stderr. It resumes interrupted downloads, unpacks under data/, and deletes the ZIP after successful extraction. Completed data is reused. Allow room for the ZIP and extracted files. Public datasets need no token; export your credential first for private ones.

The manifest command handles archives without dataset.json. It uses this dataset's baseline label mappings, weights, and LiDAR normalization; it does not recompute statistics. ZOD selects mode-specific camera, LiDAR, or fusion annotations. Six label IDs map to background, vehicle, sign, and human.

Training frames	Validation frames	Test splits	Visualization frames
1,148	573	5	8

If manifest generation reports missing files, inspect the affected entries. Missing annotations become empty masks during training. Restore them or remove the frame from the affected split before an accuracy experiment. Warnings for unused annotation folders do not affect your chosen config.

First run¶

One epoch: CLFTv2 fusion

visin-fusion run -c configs/examples/zod/clftv2/fusion.json --upload --benchmark-device cuda

Use --benchmark-device cpu on a CPU host. Outputs go to logs/zod/clftv2/fusion/. Configure Visin to upload the run, metrics, and images. One epoch is a pipeline check.

In a new terminal, activate .venv and repeat the preparation block; downloaded files are reused.

Next: Choose another model or mode. See Datasets for layout, normalization, and your own data.