Ready-to-run recipes that turn public or local datasets into Pixano-importable sources.
Every recipe produces the same thing: a source folder with a dataset.yaml manifest at
its root and one metadata.jsonl per split (JSONL v2 — see the
Importing Data reference).
Import any of them with a single command, no Python schema files needed:
pixano data import ./my_data ./<sample_folder>Requires pixano ≥ 0.8 (pip install pixano).
| Recipe | Task | Data source | Extra deps |
|---|---|---|---|
data_importation/voc |
Object detection (bbox) | Pascal VOC 2007, auto-downloaded | pillow |
data_importation/davis |
Video object segmentation (mask sidecars, tracklets) | DAVIS 2017 root, or --synthetic N (no download) |
pillow numpy (synthetic) |
data_importation/flir |
Multi-view RGB+thermal detection (image or video mode) | FLIR ADAS v2 root (manual download) | — |
data_importation/vqav2 |
Visual question answering (conversations) | HuggingFace merve/vqav2-small |
datasets pillow |
data_importation/mel |
Image–text entity linking (bbox + text spans) | Fully synthetic | pillow |
data_importation/unlabeled_images_folder |
Media-only import (no metadata at all) | Your image folder | — |
data_importation/unlabeled_videos_folder |
Frame-sequence import from raw videos | Your video folder | opencv-python |
Each generate_sample.py has a docstring with its exact usage; every script ends by printing
the import command for the folder it just produced. Use pixano data import … --dry-run to
validate a source and preview the plan without writing anything.
bash scripts/check_all.shGenerates a small sample for every offline-capable recipe and runs a dry-run and a real
import for each (into a temporary data directory). Recipes needing local dataset roots run
only when DAVIS_ROOT / FLIR_ADAS_ROOT are set; the VQAv2 recipe needs network access to
HuggingFace and is skipped otherwise.