Skip to content

Analysis on training result caused by different normalization scheme I

Jump to bottom
NoBa1anc3 edited this page May 18, 2020 · 2 revisions

Comparison

使用ImageNet标准化与公司章程数据集标准化和不使用标准化训练比较图如下，具体训练数据见附录。

附录

avg mean and std of ImageNet

Config

  • batchsize : 2
  • flip_ratio : 0
  • learning_rate : 1e-4
  • img_mean : (123.675, 116.28, 103.53)
  • img_std : (58.395, 57.12, 57.375)
  • shuffle : False
  • test_dataset_pad_mode : non-fixed

Evaluation results for bbox

Average Precision
Batch AP AP50 AP75 APs APm APl
500 0.9 3.9 0.3 nan 0.0 0.9
1000 4.1 16.3 0.7 nan 0.8 4.9
1500 5.6 19.8 1.3 nan 1.9 6.2
2000 11.5 31.1 4.6 nan 2.9 14.7
2500 14.1 36.0 6.6 nan 8.4 14.9
3000 14.0 35.8 8.0 nan 9.3 14.1
3500 14.6 35.4 8.8 nan 9.4 15.8
4000 19.9 47.0 12.6 nan 17.4 19.6
4500 18.6 45.0 10.8 nan 13.1 19.0
5000 25.1 53.9 20.7 nan 19.6 25.4
6000 25.2 54.3 18.6 nan 13.4 27.6
6500 25.6 55.2 19.5 nan 24.3 25.0
7000 29.5 62.1 22.6 nan 25.9 28.8
7500 33.0 64.1 28.0 nan 27.1 33.7
8000 33.8 66.4 30.9 nan 25.4 34.6
8500 37.1 70.5 34.8 nan 27.9 37.9
9000 34.0 70.0 27.2 nan 25.0 34.1
9500 38.8 72.3 36.8 nan 33.3 38.9
10000 34.1 69.5 28.8 nan 28.6 34.1
Average Recall
Batch AR1 AR10 AR100 ARs ARm ARl
500 0.8 3.8 5.8 nan 0.0 6.0
1000 7.4 16.3 18.9 nan 4.3 21.0
1500 7.9 15.3 17.6 nan 4.3 19.1
2000 16.0 26.6 29.0 nan 13.7 30.8
2500 18.7 30.8 32.7 nan 21.2 33.1
3000 18.1 29.4 31.8 nan 24.4 30.8
3500 18.6 31.6 34.1 nan 25.8 33.1
4000 24.4 37.4 39.8 nan 31.4 39.1
4500 20.4 31.7 34.1 nan 17.8 35.4
5000 27.8 42.0 44.3 nan 35.2 44.2
6000 26.3 40.6 42.0 nan 23.9 43.8
6500 27.0 42.6 44.7 nan 42.7 43.3
7000 29.7 44.7 46.6 nan 34.7 46.7
7500 31.6 48.2 49.8 nan 40.9 49.3
8000 30.9 48.2 50.0 nan 40.3 49.6
8500 32.9 51.7 52.8 nan 39.4 52.9
9000 30.8 49.5 50.6 nan 42.4 50.5
9500 33.4 51.9 52.7 nan 46.5 52.0
10000 30.3 48.7 50.0 nan 46.2 49.5

avg mean and std of Company Articles Dataset

Config

  • batchsize : 2
  • flip_ratio : 0
  • learning_rate : 1e-4
  • img_mean : (0.9684, 0.9683, 0.9683)
  • img_std : (0.1502, 0.1505, 0.1505)
  • shuffle : False
  • test_dataset_pad_mode : non-fixed

Evaluation results for bbox

Average Precision
Batch AP AP50 AP75 APs APm APl
500 0.3 1.4 0.0 nan 0.0 0.3
1000 2.1 8.9 0.3 nan 0.4 2.7
1500 4.3 13.7 1.7 nan 0.5 5.4
2000 13.9 33.1 8.5 nan 8.3 15.5
2500 15.7 38.4 7.5 nan 12.9 15.8
3000 7.1 21.5 1.9 nan 3.4 7.8
3500 19.4 43.0 13.8 nan 16.4 20.1
4000 25.3 53.0 22.1 nan 22.2 24.7
4500 26.3 53.6 25.2 nan 25.0 25.3
5000 28.4 57.8 24.4 nan 27.5 27.4
5500 33.1 63.9 31.6 nan 31.6 32.3
6000 14.7 34.1 11.3 nan 8.9 16.7
6500 5.5 19.1 0.8 nan 4.5 5.8
7000 32.4 68.3 21.7 nan 33.2 31.0
7500 13.1 30.0 8.5 nan 12.2 14.4
8000 12.9 37.6 2.7 nan 9.0 13.0
8500 1.5 5.2 0.2 nan 0.0 1.6
9000 17.9 39.4 13.5 nan 17.2 18.1
9500 0.0 0.0 0.0 nan 0.0 0.0
10000 12.9 28.8 8.7 nan 14.7 12.6
10500 4.1 13.6 0.2 nan 3.9 4.6
11000 0.7 3.2 0.0 nan 0.0 0.7
11500 2.3 8.1 0.4 nan 1.6 2.2
12000 0.0 0.0 0.0 nan 0.0 0.0
12500 16.6 44.1 4.9 nan 19.9 15.8
13000 34.1 73.1 22.6 nan 37.8 33.1
13500 2.1 8.6 0.1 nan 2.6 2.2
14000 55.5 86.5 68.9 nan 44.8 55.1
Average Recall
Batch AR1 AR10 AR100 ARs ARm ARl
500 0.8 2.2 2.7 nan 0.0 2.9
1000 5.9 12.7 13.7 nan 5.1 15.0
1500 7.3 17.6 19.3 nan 4.2 21.1
2000 17.4 29.7 32.0 nan 19.8 33.0
2500 19.5 34.0 36.8 nan 24.6 36.9
3000 11.5 18.6 18.9 nan 6.1 21.2
3500 22.3 36.8 38.8 nan 30.2 38.0
4000 27.7 42.7 44.6 nan 35.9 43.6
4500 27.1 42.0 44.0 nan 36.2 43.2
5000 29.0 43.6 45.3 nan 38.0 44.6
5500 31.5 47.3 48.9 nan 42.3 47.7
6000 20.6 33.6 34.4 nan 28.0 34.4
6500 6.6 10.6 10.6 nan 5.1 11.3
7000 30.2 45.4 47.4 nan 42.5 47.3
7500 17.6 23.1 23.3 nan 23.3 22.9
8000 11.1 18.6 18.6 nan 9.1 19.5
8500 1.4 2.7 2.7 nan 0.0 2.9
9000 22.1 30.5 31.0 nan 33.4 29.5
9500 0.0 0.0 0.0 nan 0.0 0.0
10000 17.7 31.1 31.9 nan 35.4 30.4
10500 6.6 7.4 7.4 nan 4.7 8.5
11000 0.9 1.8 1.8 nan 0.0 1.9
11500 2.6 3.7 3.7 nan 1.5 3.8
12000 0.0 0.0 0.0 nan 0.0 0.0
12500 20.3 30.9 31.6 nan 24.2 31.8
13000 29.7 44.1 44.5 nan 48.3 44.5
13500 3.5 3.7 3.7 nan 2.6 4.2
14000 38.8 64.6 65.0 nan 61.2 63.8

without avg mean and std

Config

  • batchsize : 2
  • flip_ratio : 0
  • learning_rate : 1e-4
  • img_mean : (0., 0., 0.)
  • img_std : (1., 1., 1.)
  • shuffle : False
  • test_dataset_pad_mode : non-fixed

Evaluation results for bbox

Average Precision
Batch AP AP50 AP75 APs APm APl
500 0.3 1.3 0.0 nan 0.0 0.3
1000 6.6 19.3 3.0 nan 2.2 7.6
1500 10.8 28.8 5.8 nan 5.8 11.8
2000 14.6 32.2 10.4 nan 10.9 15.0
2500 16.5 42.5 8.4 nan 15.9 16.8
3000 22.2 51.5 13.9 nan 20.8 21.8
3500 19.5 46.3 11.1 nan 14.2 19.5
4000 27.4 57.5 22.9 nan 24.7 26.4
4500 27.5 56.5 23.1 nan 28.0 26.3
5000 36.0 70.7 30.6 nan 28.9 36.1
5500 38.4 72.7 38.8 nan 34.1 37.2
6000 38.3 74.7 34.7 nan 21.4 40.8
6500 34.8 71.3 30.4 nan 24.7 34.7
7000 33.8 73.2 22.0 nan 25.2 32.9
7500 47.8 84.1 51.2 nan 37.5 47.5
8000 4.0 12.4 0.9 nan 1.2 4.4
8500 29.4 65.9 19.6 nan 32.8 27.5
9000 44.8 82.5 45.5 nan 37.1 43.7
9500 14.7 35.8 8.5 nan 9.7 15.7
10000 51.0 88.5 52.1 nan 41.6 50.6
Average Recall
Batch AR1 AR10 AR100 ARs ARm ARl
500 0.4 2.1 2.1 nan 0.0 2.2
1000 11.1 21.0 23.1 nan 9.1 24.6
1500 15.0 27.2 29.1 nan 16.2 29.7
2000 14.8 31.5 33.7 nan 20.8 34.4
2500 20.0 33.8 35.9 nan 23.9 35.9
3000 24.1 38.6 40.7 nan 36.9 39.4
3500 19.7 32.7 34.3 nan 20.9 34.5
4000 26.8 41.8 43.1 nan 33.4 42.6
4500 26.1 40.9 43.0 nan 35.9 41.8
5000 30.9 49.0 50.1 nan 46.7 49.4
5500 32.4 52.0 53.2 nan 48.6 51.9
6000 31.0 50.1 51.0 nan 32.6 51.7
6500 28.8 48.3 49.3 nan 42.2 49.0
7000 28.1 43.7 43.9 nan 27.8 44.3
7500 35.9 59.1 59.7 nan 51.5 58.9
8000 3.9 7.2 7.2 nan 1.7 7.6
8500 27.5 45.3 46.2 nan 42.4 45.1
9000 34.8 56.7 57.4 nan 56.6 56.1
9500 14.2 21.0 21.0 nan 10.1 22.5
10000 37.0 59.9 60.3 nan 53.8 60.0
Clone this wiki locally