צורות 3d

3dshapes הוא מערך נתונים של צורות תלת-ממדיות שנוצרו באופן פרוצדורלי מ-6 גורמים סמויים בלתי תלויים של אמת קרקע. גורמים אלה הם צבע הרצפה, צבע הקיר, צבע האובייקט , קנה המידה , הצורה והכיוון .

כל השילובים האפשריים של הסמויים הללו קיימים בדיוק פעם אחת, ומייצרים N = 480000 תמונות בסך הכל.

ערכי גורמים סמויים

  • גוון רצפה: 10 ערכים ברווח ליניארי ב-[0, 1]
  • גוון קיר: 10 ערכים ברווח ליניארי ב-[0, 1]
  • גוון אובייקט: 10 ערכים ברווח ליניארי ב-[0, 1]
  • קנה מידה: 8 ערכים ברווח ליניארי ב-[0, 1]
  • צורה: 4 ערכים ב-[0, 1, 2, 3]
  • כיוון: 15 ערכים ברווח ליניארי ב-[-30, 30]

שינינו לטנט אחד בכל פעם (החל מכיוון, ואז צורה וכו'), ואחסנו ברצף את התמונות בסדר קבוע במערך images . הערכים המתאימים של הגורמים מאוחסנים באותו סדר במערך labels .

לְפַצֵל דוגמאות
'train' 480,000
  • מבנה תכונה :
FeaturesDict({
    'image': Image(shape=(64, 64, 3), dtype=uint8),
    'label_floor_hue': ClassLabel(shape=(), dtype=int64, num_classes=10),
    'label_object_hue': ClassLabel(shape=(), dtype=int64, num_classes=10),
    'label_orientation': ClassLabel(shape=(), dtype=int64, num_classes=15),
    'label_scale': ClassLabel(shape=(), dtype=int64, num_classes=8),
    'label_shape': ClassLabel(shape=(), dtype=int64, num_classes=4),
    'label_wall_hue': ClassLabel(shape=(), dtype=int64, num_classes=10),
    'value_floor_hue': float32,
    'value_object_hue': float32,
    'value_orientation': float32,
    'value_scale': float32,
    'value_shape': float32,
    'value_wall_hue': float32,
})
  • תיעוד תכונה :
תכונה מעמד צוּרָה Dtype תיאור
FeaturesDict
תמונה תמונה (64, 64, 3) uint8
תווית_קומה_גוון ClassLabel int64
label_object_hue ClassLabel int64
label_orientation ClassLabel int64
label_scale ClassLabel int64
label_shape ClassLabel int64
label_wall_hue ClassLabel int64
ערך_גוון_קומה מוֹתֵחַ לצוף32
value_object_hue מוֹתֵחַ לצוף32
ערך_אוריינטציה מוֹתֵחַ לצוף32
value_scale מוֹתֵחַ לצוף32
צורת_ערך מוֹתֵחַ לצוף32
value_wall_hue מוֹתֵחַ לצוף32

רְאִיָה

  • ציטוט :
@misc{3dshapes18,
  title={3D Shapes Dataset},
  author={Burgess, Chris and Kim, Hyunjik},
  howpublished={https://github.com/deepmind/3dshapes-dataset/},
  year={2018}
}