# Previews for the Dataset Viewer These Parquet files exist only to power the Hugging Face Dataset Viewer on the repository page. **Do not train on these.** They hold downscaled JPEG previews, the split, per-class pixel percentages, and the paths of the full-resolution files. | Config | Files | One row per | Images | |---|---|---|---| | `default` | `train`, `validation`, `test` | hand-labeled frame pair in `Data/` (951) | RGB and overlay at 1344 x 756, IR at 640 x 360 | | `pseudo` | `pseudo_*` | automatically labeled frame in `PseudoLabeling/PseudoDataset10s/` (2,214) | RGB and overlay at 1008 x 567, IR at 640 x 360 | Both configs use the paper's Graminoid-held-out partition (`Data/splits_graminoid.csv`): Graminoid is the test set, Canopy provides train and validation. Pseudo frames take the split of their anchor frames. The `split` column of `pseudo_samples.csv`, which holds Canopy out instead, is kept as `split_canopy_held_out`.