Add and Prepare Data

Link to this sectionWhat Platform accepts#

Platform datasets accept uploads or connected cloud storage — see the data docs for the full file-format matrix, and the data collection and annotation guide plus preprocessing annotated data guide for the upstream methodology:

Storage is content-addressable (XXH3-128 hashing), so re-uploading the same image is free — Platform stores it once and references it everywhere.

Cloud-backed datasets keep originals in customer storage and stream pinned revisions when needed. They currently exclude auto-annotation, clustering, cloning, and immutable version snapshots; use an uploaded dataset when the course requires those features.

Link to this sectionVideo handling#

Video uploads are simple by design: Platform extracts frames at a fixed 1 fps, capping at 100 frames per video (so a 30-fps, 30-minute clip becomes the first 100 seconds — not 54 000 frames). For longer or more selective sampling, pre-sample the source yourself and upload the resulting frames.

   30-minute, 30-fps source video
   ────────────────────────────────────────────────────►  54,000 frames
   Platform extraction (1 fps, max 100):
   ▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪
                                                          100 frames (first 100s)

For event-driven sampling — only keep frames where something interesting is happening — run a pretrained Ultralytics YOLO model over the source first and upload the survivors:

   plain 1 fps sampling           event-driven sampling (pre-filter with YOLO)
   ┌─────────────────────────┐    ┌─────────────────────────┐
   │ ▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪▪ │    │ ▪      ▪▪▪▪▪      ▪▪▪▪▪ │
   │ (empty corridor)        │    │ (sparse)  ↑ (rich)      │
   │                         │    │      target visible     │
   └─────────────────────────┘    └─────────────────────────┘
   1800 frames, ~50 with target   ~400 frames, ~80 with target

ffmpeg -i source.mp4 -vf fps=1 frames/%06d.jpg

Then upload the frames/ directory or zip it into a dataset archive — Platform's content-addressable storage will dedupe identical frames automatically.

Link to this sectionPicking a task type#

A dataset's task type is set on creation and decides which annotation tools appear:

You can change the task later from the dataset header, but incompatible annotations stop displaying — pick the one that matches your downstream model.

Link to this sectionInspect what you uploaded#

Once images are processed, the dataset's Charts tab shows automatic statistics:

• Train / val / test split distribution.
• Top class distribution (donut chart).
• Image width / height histograms.
• Annotation location heatmap.
• Image dimension scatter with aspect-ratio guide lines.

Skim these before you start labeling — they catch class imbalance, undersized images, and weird aspect ratios early.

Link to this sectionCloning a public dataset to start#

If you'd rather not start from scratch, the Explore page lists official Ultralytics and community datasets you can clone with one click and customize. Standard YOLO benchmarks like COCO and Open Images V7 are also drop-in options for any local training that uses the ul:// URI scheme — see the datasets catalog for the full list. The data collection and annotation guide goes deeper on filling rare-class gaps and avoiding obvious sampling pitfalls.