• Train the model (Scenario 1)

    • batch size 8
    • Val Freq: 10
    • LR 1e-5
    • Epoch = 1

    conf_1_test_confmatrix.png

  • Scenario 2: Split Noise. Below 40%(find the threshold)

    • Gloria Poisson

      • 5

    • Nicolo Salt and Pepper

      • 0.1 (10%) = 0.18 acc
      • 0.05 (5%) = 0.27
      • 0.02 (2%) = 0.43

      Untitled

      Untitled

      • 0.01 (1%) = 0.54
      • 0.001 (0.1%) = 0.88

      Untitled

    Untitled

    • Daniel Gaussian

  • 0,15

  • 0.07 std = 0.9 acc

    conf_2_test_confmatrix_good.png

    conf_2_test_images_good.png

  • Scenario 3: use the same parameters but with noise in the training and hope the acc is higher (above 80%)

    • Nicolò: Salt and pepper (0.02 noise = 0.95 test acc)

    Untitled

    Untitled

    • Daniel: Gaussian

      conf_3_test_images_37_gaus.png

      conf_3_test_confmatrix_gaus.png

    • Gloria: Poisson

      • lambda 5

      Untitled

      Untitled

      Histogram

      Original image

      Untitled

      Augmented image

      Untitled