MNIST Double Digits Datasets Classification and Detectection

Datasets:

• 64x64 size images containing two digits
• digits range [0, 9]
• digits size (bounding box size) 28x28

Algorithm used:

• Classes: classification

  • two digits range from 0 to 9
  • model returned labels range from 0 to 19
  • reshape the labels into shape [N, 2]
  • each n data contain two class that the two digits belongs to

• Bounding Boxes: Classification

  • two digits' position ranges from 0 to 37 ((64+1)-28)
  • model returned labels range from 0 to 148 (4*37)
  • reshape the labels in to shape [N, 2, 4]
  • each n data contain [x1, y1, x1+28, y1+28], [x2, y2, x2+28, y2+28] where x and y represents the axis and 1 and 2 represent first and second digits

Model:

    

Result:

• Validation datasets with model.pt.172:
  • Classification accuracy: 99.63
  • Detection IoU: 95.17
  • Valid time (GPU): ≈23.28
  • Valid time (CPU): ≈418