cifar10_h

  • Descrizione :

Una versione rietichettata del set di test CIFAR-10 con etichette soft provenienti da veri annotatori umani. Per ogni coppia (immagine, etichetta) nel set di test CIFAR-10 originale, fornisce diverse etichette aggiuntive fornite da annotatori umani reali, oltre alla soft-label media. Il set di addestramento è identico a quello del set di dati originale.

Diviso Esempi
'test' 10.000
'train' 50.000
  • Struttura delle caratteristiche :
FeaturesDict({
    'annotator_ids': Sequence(Scalar(shape=(), dtype=int32)),
    'human_labels': Sequence(ClassLabel(shape=(), dtype=int64, num_classes=10)),
    'id': Text(shape=(), dtype=string),
    'image': Image(shape=(32, 32, 3), dtype=uint8),
    'label': ClassLabel(shape=(), dtype=int64, num_classes=10),
    'reaction_times': Sequence(Scalar(shape=(), dtype=float32)),
    'soft_label': Tensor(shape=(10,), dtype=float32),
    'trial_indices': Sequence(Scalar(shape=(), dtype=int32)),
})
  • Documentazione delle funzionalità :
Caratteristica Classe Forma Tipo D Descrizione
CaratteristicheDict
annotator_ids Sequenza (scalare) (Nessuno,) int32
etichette_umane Sequenza(ClassLabel) (Nessuno,) int64
id Testo corda
Immagine Immagine (32, 32, 3) uint8
etichetta ClassLabel int64
tempi_di_reazione Sequenza (scalare) (Nessuno,) float32
soft_label Tensore (10,) float32
prova_indices Sequenza (scalare) (Nessuno,) int32

Visualizzazione

  • Citazione :
@inproceedings{wei2022learning,
  title={Human uncertainty makes classification more robust},
  author={Joshua C. Peterson and Ruairidh M. Battleday and Thomas L. Griffiths
  and Olga Russakovsky},
  booktitle={IEEE International Conference on Computer Vision and Pattern
  Recognition (CVPR)},
  year={2019}
}